IAES Inter national J our nal of Robotics and A utomation (IJRA) V ol. 14, No. 4, December 2025, pp. 528 ∼ 538 ISSN: 2722-2586, DOI: 10.11591/ijra.v14i4.pp528-538 ❒ 528 Edge-awar e distilled segmentation with pseudo-label r enement f or autonomous dri ving per ception No v elio Putra Indarto, Oskar Natan, Andi Dharmawan Department of Computer Science and Electronics, F aculty of Mathematics and Natural Sciences, Uni v ersitas Gadjah Mada, Y ogyakarta, Indonesia Article Inf o Article history: Recei v ed Jul 1, 2025 Re vised Oct 24, 2025 Accepted No v 11, 2025 K eyw ords: DenseCRF Edge computing Ef cientNet Entrop y masking Kno wledge distillation Pseudo-label renement Semantic se gmentation ABSTRA CT Achie ving precise semanti c se gmentation is essential for enabling real-time perception in autonomous systems, yet leading approaches typically require substantial annotated data and po werful hardw are, restricting their use on de- vices with limited resources. This w ork introduces an ef cient se gmentation frame w ork that inte grat es pseudo-label renement, kno wledge distillation, and entrop y-based condence lter ing to train compact student netw orks suitable for edge deplo yment. High-quality pseudo-label s are rst produced by a rob ust teacher netw ork, then further impro v ed using a dense conditional random eld to boost spatial consistenc y . An entrop y-based selection mechanism remo v es unreliable predictions, ensuring that only the mos t trustw orth y labels guide the student model’ s training. The use of kno wledge distillation ef fecti v ely transfers detailed semantic understanding from the teacher to the student, enhancing accu- rac y without added computational o v erhead. Experimental results with multiple Ef cientNet backbones re v eal that this pipeline impro v es se gmentation accurac y and output clarity , while also supporting real-time or near real-time inference on CPUs with limited processing po wer . Extensi v e ablation and qualit ati v e stud- ies further conrm the method’ s rob ustness and e xibility for real-w orld edge applications. This is an open access article under the CC BY -SA license . Corresponding A uthor: Oskar Natan Department of Computer Science and Electronics, F aculty of Mathematics and Natural Sciences, Uni v ersitas Gadjah Mada Sleman Re genc y , Special Re gion of Y ogyakarta, Indonesia Email: oskarnatan@ugm.ac.id 1. INTR ODUCTION The widespread adoption of autonomous v ehicles depends on rob ust, real-time perception systems such as semantic se gmentation, which are essential for safe na vig ation in comple x en vironments [1]–[5]. While state-of-the-art deep learning models lik e DeepLabV3+ [6], HRNet [7], and transformer -based architecture lik e Se gF ormer [8] achie v e remarkable accurac y , the y often require signicant computational resources, making them dif cult to deplo y on embedded de vice s or consumer -grade v ehicles with limited hardw are capabilities [9]–[11]. T o address these constraints, researc hers ha v e de v eloped v arious strate gies for model compression and acceleration, including netw ork pruning [12], parameter quantization [13], and the creation of lightweight architectures lik e Enet [14], MobileNet [15], and Ef cientNet [16]. Ho we v er , lightweight models alone often struggle to maintain the detailed accurac y required for real-w orld, urban en vironments. Kno wledge distillation (KD) has emer ged as a po werful approach for bridging the g ap between model ef cienc y and performance. By J ournal homepage: http://ijr a.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
IAES Int J Rob & Autom ISSN: 2722-2586 ❒ 529 transferring kno wledge from a lar ge, accurate teacher model to a smaller student, KD enables compact models to approach the accurac y of their lar ger counterparts with much lo wer resource requirements [17]. F or dense prediction tasks lik e semantic se gmentation, KD has been e xtended through feature imitation, attention transfer , and boundary-a w are distillation [18]. Despite these adv ance s, most KD methods still rely on the a v ailability of annotated data, which remains a signicant bottleneck due to the high cost and labor required for pix el-le v el labeling, especially in rapidly e v olving urban scenes [18], [19]. This challenge has moti v ated increased interest in self-supervised, semi-supervised, and teacher -student learning frame w orks that can le v erage lar ge amounts of unlabeled data [20]. A k e y technique in this conte xt is pseudo-labeling, where pro visional labels are assigned to unlabe led samples based on the model’ s condent predictions and then used to further train the netw ork [21]. Ho we v er , na ¨ ıv e pseudo-labeling can introduce noise and conrmation bias if unreliable predictions are included. T o mitig ate this, recent w orks use condence thresholding, uncertainty estimation, and ensemble methods to lter ambiguous predictions [22], [23]. Spatial renement methods lik e dense conditional random elds (DenseCRF) further enhance pseudo-label quality by enforcing consistenc y among neighboring pix els [24]. Additionally , entrop y-based masking strate gies allo w models to dynamically select reliable re gions for supervision, reduc- ing the propag ation of errors and impro ving o v erall training stability [25], [26]. Although recent adv ances in semi-supervised and self-supervised learning ha v e made signicant strides in reducing reliance on lar ge la- beled datasets for semantic se gmentation, the majori ty of e xisting methods still require some le v el of manual annotation, whether for pretraining, v alidation, or h yperparameter tuning [27]. Moreo v er , ongoing research predominantly na vig ates controlled e xperimental en vironments where benchmarks may not ef fecti v ely reect the real-w orld comple xities such as v ariable lighting, occlusions, or limited computational ca pacities. As a consequence, the real-w orld adoption of these approaches is oft en limited, especially for applications requiring ef cient, rob ust, and easily deplo yable se gmentation on hardw are with strict resource constraints. In f act, only a small number of studies ha v e specically addressed the practical challenges of deplo ying se gmentation mod- els for real-time inference on edge de vices with limited memory and processing po wer , such as those used in embedded robotics and autonomous v ehicles that highlighting an ongoing need for streamlined solutions that do not depend on e xtensi v e annotation or cumbersome training procedures [28], [29]. In this w ork, we propose a unied, edge-a w are, and fully self-supervised semantic s e gm entation pipeline tailored for urban autonomous dri ving, which eliminates the need for manual annotation at an y stage. Our approach combines a transformer -based teacher for pseudo-label generation, DenseCRF for boundary renement, entrop y-based condence masking, a lightweight Ef cientNet-U-Net student, and a streamlined logit-based KD loss. W e v alidate our frame w ork using a real-w orld dataset from Uni v ersitas Gadjah Mada (UGM), Indonesia, and demonstrate ef fecti v e real-time deplo yment on simulated embedded hardw are. Our contrib utions include: i) a fully self-supervised pipeline for urban se gmentation without manual labels, ii) a thorough analysis of pseudo-label renement, kno wledge distillation, and condence thresholding, and iii) de- plo yment and benchmarking on resource limited CPU or simulated edge-de vice conditions. Finally , we discuss the limi tations and future directions for annotation-free se gmentation, particularly re g arding e v aluation without ground truth in practical scenarios. 2. METHOD Our proposed method addresses semantic se gmentation for autonomous dri ving in resource-constrained urban en vironments, entirely without ground truth annotation. The pipeline consists of three main components: a teacher model for pseudo-label generation, a renement step using DenseCRF to enhance label quality , and a student model trained with kno wledge distillation and entrop y-based condence masking. The o v erall archi- tecture is illustrated in Figure 1. 2.1. T eacher model and pseudo-Label generation W e chose Se gF ormer -B5 [8] as our teacher model due to its e xceptional se gmentation perfor mance on urban-scene benchmarks such as Cityscapes [30]. Se gF ormer is the one of SO T A models that inte grates T ransformer -based attention mechanisms, ef fecti v ely capturing global conte xtual information, which is critical for accurately se gmenting cro wded urban en vironments. The pre-trained teacher model generates predictions for the unlabeled images in our dataset. Specically , we emplo y the teacher model as a ”pseudo annotator , ” generating preliminary se gmentation masks (pseudo-labels). The teacher outputs softmax probability maps Edg e-awar e distilled semantic se gmentation with pseudo-label r enement ... (No velio Putr a Indarto) Evaluation Warning : The document was created with Spire.PDF for Python.
530 ❒ ISSN: 2722-2586 P ∈ [0 , 1] H × W × C , where C denotes the number of classes. The hard pseudo-labels Y at pix el ( i, j ) is deter - mined by selecting the class with maximum probability . Y i,j = arg max c P i,j ,c (1) Figure 1. Complete pipeline of our method 2.2. Pseudo-label r enement with DenseCRF While the teacher’ s ps eudo-labels Y pro vide strong supervisory s ignals, the y may contain spatial noise and misaligned boundaries. T o address this, we apply Dens eCRF [24] to each pseudo-label mask. DenseCRF models both unary potentials (from the teacher’ s predictions) and pairwise potentials that encourage spatial and appearance consistenc y among neighboring pix els. The ener gy function for the CRF is dened as (2): E ( Y , I ) = X i ψ u ( Y i | I ) + X i<j ψ v ( Y i , Y j | I ) (2) where Y be the v ector formed by the Y u and Y v pseudo-labels, where Y u represents the unary potentials and Y v represents the pairwise potentials. This post-processing step yields rened pseudo-labels ˜ Y with impro v ed boundary accurac y and spatial coherence serv e as supervision for training the student model. 2.3. Entr opy-based condence masking Ev en after renement, pseudo-labels may contain re gions with high uncertainty , typically occurring at class boundaries, object occlusions, or in visually ambiguous areas . T raining directly on these uncertain labels may cause the student model to learn erroneous mappings, ultimately de grading generalization and se g- mentation accurac y [ 3 1] . T o mitig ate this, we emplo y entrop y-based condence masking, a widely recognized approach i n semi-supervised and pseudo-label learning [32]. The entrop y of the teacher’ s predict ed probability v ector for pix el (i,j) is dened as: H ( i, j ) = − X c P i,j ,c log P i,j ,c (3) Entrop y , as a measure of uncertainty , reaches its maxim um when the model is maximally uncertain and is minimized when the prediction is highly condent. W e then apply a threshold τ to lter out pix els with high entrop y , retaining only those with lo w uncertainty . The condence mask M is dened as: M ( i, j ) = ( 1 if H ( i, j ) < τ 0 otherwise (4) IAES Int J Rob & Autom, V ol. 14, No. 4, December 2025: 528-538 Evaluation Warning : The document was created with Spire.PDF for Python.
IAES Int J Rob & Autom ISSN: 2722-2586 ❒ 531 where τ is a h yperparameter that can be tuned based on the dataset and model perform ance. Only pix els with M ( i, j ) = 1 are considered in loss computation, ensuring that the student model learns only from high- condence pseudo-labels. 2.4. Student model training with kno wledge distillation The student model architecture is based on Ef cientNet [16] as encoder , selected for its strong performance-to-ef cienc y ratio in edge and embedded applications. T o enable spatially accurate se gmentation, we pair this encoder with a U-Net–style decoder [33], which uses skip connections to preserv e ne-grained features across multiple resolutions. The student is trained using a composite loss function that inte grates tw o complementary objecti v es: a se gmentation loss to enforce agreement with high-condence, rened pseudo- labels, and a kno wledge distillation loss to transfer richer semantic structure from the teacher . The total training loss is formulated as: L = L seg + λL k d (5) where L seg is the cross-entrop y loss computed only o v er pix els where M ( i, j ) = 1 , λ is a h yperparameter balancing the tw o losses, and L k d is the kno wledge distillation loss. 2.4.1. Segmentation loss The se gmentation loss L seg is dened as the sum of pix el-wise cross-entrop y and dice loss between the student’ s predicted probability map Q and the rened pseudo-labels ˜ Y , weighted by the condence mask M : L seg = L ce + L dice (6) L ce = − X i,j M ( i, j ) ˜ Y ( i, j ) log Q ( i, j ) + (1 − ˜ Y ( i, j )) log (1 − Q ( i, j )) (7) where S ( i, j , c ) is the student’ s predicted probabil ity for class c at pix el ( i, j ) . L dice = 1 − 2 P i,j M ( i, j ) ˜ Y ( i, j ) Q ( i, j ) P i,j M ( i, j ) ˜ Y ( i, j ) + P i,j M ( i, j ) Q ( i, j ) (8) The Dice loss is particularly ef fecti v e for handling cl ass imbalance and impro ving the accurac y of object bound- aries, which are critical for safety in autonomous dri ving scenarios. 2.4.2. Kno wledge distillation loss T o further enhance the student’ s learning, we apply kno wledge distillation (KD) in the form of K ullback- Leibler (KL) di v er gence between the temperature-scaled softmax outputs of the teacher and student [17]: L k d = 1 N X i,j M i,j X c P i,j ,c log P i,j ,c Q i,j ,c (9) where P i,j ,c is the teacher’ s softmax output, and Q i,j ,c is the student’ s softmax output. W e also apply a tem- perature scaling f actor T to soften the logits before computing the KL di v er gence. This loss encourages the student to mimic the teacher’ s distrib ution o v er classes, capturing richer semantic information be yond simple class labels. 3. RESUL TS AND DISCUSSION W e e v aluated the model’ s performance and reliability from se v eral perspecti v es, including stabil ity of the teacher model, ef fecti v eness of loss function, qualitati v e outcomes, comparati v e analysis with baseline methods, computational ef cienc y , and real-w orld applicability through of ine and online testing. T o better understand the impact of each component, we conducted a qualitati v e analysis of the teacher model’ s pseudo- labels before and after renement, as well as an ablation study on the student model’ s performance with and without kno wledge distillation and entrop y-based condence masking across v arious Ef cientNet backbones (B0, B3, B5, B7). Edg e-awar e distilled semantic se gmentation with pseudo-label r enement ... (No velio Putr a Indarto) Evaluation Warning : The document was created with Spire.PDF for Python.
532 ❒ ISSN: 2722-2586 3.1. Ev aluation of pseudo-label r enement T o thoroughly assess the impact of pseudo-label renement, we conduct a series of qualitati v e analy- ses comparing the teacher’ s initial pseudo-labels and the results after DenseCRF post-process ing. As sho wn in Figure 2. In this gure, each ro w corresponds to a dif ferent scene, with columns representing Figure 2(a) the original input image, Figure 2(b) the teacher’ s ra w pseudo-labels, Figure 2(c) the pseudo-labels after DenseCRF renement, and Figure 2(d) the entrop y map of the teacher’ s softmax predictions. The yello w box es highlight k e y re gions of interest for closer comparison. The entrop y maps clearly re v eal areas of high predicti v e uncer - tainty , indicated by brighter colors. These re gions t end to cluster along object boundaries, in zones with hea vy occlusion, or within visually comple x backgrounds—precisely the spots where the teacher model is least cer - tain about class assignments. When we e xamine the corresponding pseudo-labels before and after renement, it becomes e vident that DenseCRF has its most pronounced impact in these high-entrop y areas. Specically , ambiguous or isolated pix els near boundaries are frequently “smoothed out, ” with the renement step encour - aging these pix els to adopt the most pre v alent local class. As a result, the rened pseudo-labels display cleaner , more continuous object shapes and boundaries, impro ving their suitability as supervision for student training. Despite these impro v ements, some challenging high-entrop y re gions persist, as seen in the residual label noise or ambiguous se gmentation outcomes e v en after renement—particularly in areas where object boundaries are especially comple x or where the input image is highly cluttered. These persistent challenges underscore the limitations of spatial renement alone and suggest that further adv ances in label denoising or uncertainty modeling may be necessary for e v en greater se gmentation delity . (a) (b) (c) (d) Figure 2. Qualitati v e analysis of pseudo-label renement (a) input image, (b) teacher’ s pseudo-labels, (c) rened pseudo-labels after DenseCRF , and (d) entrop y map of teacher’ s predictions Comparing student performance when trained on pseudo-labels before and after applying DenseCRF . As summarized in T able 1, the use of rened pseudo-labels consistently results in impro v ed mIoU and F1 Score across all models. F or instance, Ef cientNet-B0’ s mIoU impro v es from 81.89% to 82.57%, and its F1 Score from 89.97% to 90.37% with renement, while the lar gest e v aluated model Ef cientNet-B7 achie v es the best IAES Int J Rob & Autom, V ol. 14, No. 4, December 2025: 528-538 Evaluation Warning : The document was created with Spire.PDF for Python.
IAES Int J Rob & Autom ISSN: 2722-2586 ❒ 533 o v erall metrics with 83.21% mIoU and 90.76% F1 after renement. These g ains highlight ho w the renement process enhances not just the visual clarity of the labels, b ut also leads to more ef fecti v e model training and generalization.Ne v ertheless, renement is not uni v ersally benecial in e v ery scenario. As illustrated by the e xamples in Figure 3, with Figure 3(a) sho ws the original input image, Figure 3(b) sho ws the teacher’ s ra w pseudo-labels, and Figure 3(c) sho ws the pseudo-labels after DenseCRF renement, there are cases where DenseCRF renement f ails to preserv e small or uncommon objects. In the highlighted re gions, rare classes such as traf c poles or signs are often subsumed by the majority class, particularly when these objects occup y a small number of pix els or are situated in visually com ple x or occluded areas. This outcome stems from the local conte xt-dri v en nature of DenseCRF , which can inadv ertently erase minority details in f a v or of smoother se gmentations. T able 1. Enhanced performance of student model with and without renement Backbone W ithout Renement W ith Renement mIoU F1 Score mIoU F1 Score Ef cientNet-B0 81.89 89.97 82.57 90.37 Ef cientNet-B3 81.52 89.74 82.75 90.49 Ef cientNet-B5 82.24 90.18 82.79 90.50 Ef cientNet-B7 82.70 90.45 83.21 90.76 (a) (b) (c) Figure 3. F ailure cases where renement absorbed rare or small objects (a) input image, (b) teacher’ s pseudo-labels, and (c) rened pseudo-labels after DenseCRF 3.2. Model perf ormance e v aluations T o rigorously assess our se gmentation pipeline, we systematically e v aluated the impact of kno wledge distillation (KD) and entrop y-based masking on se gmentation performance using rened pseudo-labels, focus- ing on mean Intersection o v er Union (mIoU) and F1 score as primary metrics. As summari zed in T able 2, in- corporating consistently impro v ed both mIoU and F1 scores across all backbone v ariants. F or Ef cientNet-B0, the smallest and most resource-ef cient model, there is a clear syner gy when combining kno wledge distilla- tion with entrop y-based ltering sho ws that the highest mIoU (82.90%) and F1 Score (90.58%) are achie v ed when the student is guided by the teacher and simultaneously shielded from lo w-condence pseudo-labels at an entrop y threshold of 0.45. This suggests that, for lightweight models, restricting learning to condent predic- tions is essential for pre v enting o v ertting to noise, while kno wledge distillation pro vides the semantic richness necessary to approach teacher -le v el accurac y e v en in edge-limited settings. The narrati v e is more comple x with Ef cientNet-B3 and B5. Notably , kno wledge distillation alone, Edg e-awar e distilled semantic se gmentation with pseudo-label r enement ... (No velio Putr a Indarto) Evaluation Warning : The document was created with Spire.PDF for Python.
534 ❒ ISSN: 2722-2586 without an y entrop y masking, produces little t o no impro v ement o v er the baseline, indicating that these mid- sized models cannot reliably e xtract v alue from teacher supervision in the presence of label noise. Ho we v er , when kno wledge distillation is paired with increasing entrop y thresholds, both B3 and B5 display a mark ed increase in performance, peaking at 82.95% mIoU and 90.61% F1 for B3 at ET 0.45, and 83.00% mIoU and 90.64% F1 for B5 at ET 0.35. These results highlight that, for mid-capacity architectures, the combination of rob ust, high-condence pseudo-labels and teacher -dri v en semantic guidance is k e y to achie ving optimal results. This also underscores the sensiti vity of these models to the quality of supervision that too much noisy or ambiguous data can obscure the benets of kno wledge transfer . T able 2. Model performance with and without KD and ET Backbone ET W ithout KD W ith KD mIoU F1 Score mIoU F1 Score Ef cientNet-B0 - 82,57 90,37 82,86 90,55 0.05 81,52 89,74 82,72 90,47 0.15 81,92 89,98 82,71 90,45 0.25 82,27 90,19 82,68 90,44 0.35 82,02 90,04 82,66 90,43 0.45 82,16 90,13 82,90 90,58 Ef cientNet-B3 - 82,75 90,49 82,37 90,25 0.05 82,37 90,25 82,75 90,49 0.15 82,17 90,14 82,23 90,17 0.25 82,56 90,37 82,63 90,41 0.35 82,65 90,42 82,83 90,53 0.45 82,81 90,52 82,95 90,61 Ef cientNet-B5 - 82,79 90,50 82,59 90,39 0.05 82,23 90,17 82,90 90,57 0.15 82,58 90,38 82,95 90,60 0.25 82,06 90,06 82,95 90,61 0.35 82,49 90,32 83,00 90,64 0.45 82,55 90,36 82,91 90,58 Ef cientNet-B7 - 83,21 90,76 83,88 91,16 0.05 82,66 90,43 83,61 91,00 0.15 82,99 90,36 83,71 91,06 0.25 83,47 90,91 83,78 91,10 0.35 83,10 90,69 83,66 91,03 0.45 83,27 90,79 83,53 90,95 The impact of entrop y-based condence masking, ho we v er , depended strongly on model capacity . F or smaller backbones such as B0 and B3, entrop y masking further impro v ed se gmentation accurac y by ltering out high-uncertainty pseudo-labels, thus reducing the risk of o v ertting to noise, a trend that w as most e vident at moderate threshold v alues. This suggests that lightweight models, being more vulnerable to label noise, benet from selecti v e training on reliable supervision. In contrast, a dif ferent pattern emer ged with the high-capacity Ef cientNet-B5. Omitting entrop y masking and utilizing all pseudo-labels, including those with greater uncer - tainty , led to the highest se gmentation metrics. This nding indicates that lar ger models can tolerate or e v en benet from e xposure to ambiguous re gions, lik ely due to their superior ability to e xtract informati v e signals and generalize from a more di v erse set of training e xamples. F or Ef cientNet-B7, the lar gest backbone e v aluated, a dif ferent trend emer ges. The model achie v es its best performance (83.88% mIoU and 91.16% F1) with kno wledge distillation alone and no entrop y threshold- ing, and sustains high metrics e v en as entrop y thresholds are v aried. This resilience suggests that high-capacity models are equipped to handle the inherent noise in pseudo-labels, capitalizing on their greater representa- tional po wer and capacity to learn from a di v erse and e v en ambiguous set of training signals. F or B7, addi- tional ltering may actually reduce the benecial e xposure to challenging, informati v e re gions that contrib ute to ne-grained se gmentation accurac y . T ak en together , these results re v eal an important interplay between model capacity , supervision qual- ity , and deplo yment considerations. While kno wledge distillation is a uni v ersally v aluable tool for narro wing the performance g ap between student and teacher models, b ut its ef fecti v eness is highly dependent on ho w supervision is deli v ered and the underlying capacity of the model. Applying entrop y-based masking can sig- nicantly enhance performance for smaller models by ensuring the y learn from high-condence signals, b ut IAES Int J Rob & Autom, V ol. 14, No. 4, December 2025: 528-538 Evaluation Warning : The document was created with Spire.PDF for Python.
IAES Int J Rob & Autom ISSN: 2722-2586 ❒ 535 may not be necessary or e v en benecial for lar ger architectures that can handle more noise. This suggests that the decision to apply entrop y-based masking should be guided by the backbone’ s representational po wer and the intended resource constraints of the deplo yment en vironment. Ov erall, our e xperiments demonstrate that the pipeline is both e xible and ef fecti v e, of fering strong se gmentation accurac y and ef cient inference for edge-a w are applications. 3.3. Edge-awar e e v aluation T o e v aluate the suitability of our se gmentation pipeline for real-w orld edge deplo yment, we cond uc ted an e xtensi v e resource-a w areness study by simulating inference on a CPU. W e use AMD Ryzen Threadripper PR O 7965WX CPU with limited 1-4 cores congurations. F or each se tting, we measured the student model’ s computational requirements and inference ef cienc y using the follo wing metrics: a v erage inference time per image, frames per second (FPS), RAM usage, and GFLOPS. Results are summarized in T able 3. T able 3. Edge-A w are Ev aluation of Student Models with V arying Cores Backbone Metric 1 Core 2 Cores 3 Cores 4 Cores Ef cientNet-B0 Size (MB): 68.95 P arameters (M): 5.80 Inferen T ime (ms) 63.93 ± 2.03 38.02 ± 2.36 28.36 ± 1.72 23.83 ± 1.05 FPS 15.66 ± 0.48 26.39 ± 1.54 35.37 ± 1.94 42.04 ± 1.81 RAM (MB) 1015.11 ± 41.79 1003.92 ± 37.37 1008.54 ± 33.75 1014.72 ± 35.47 GFLOPS 98.45 ± 3.01 165.96 ± 9.68 222.43 ± 12.20 264.36 ± 11.38 Ef cientNet-B3 Size (MB): 146.97 P arameters (M): 12.48 Inferen T ime (ms) 89.86 ± 2.05 55.03 ± 1.93 42.21 ± 1.50 34.76 ± 1.22 FPS 11.13 ± 0.26 18.19 ± 0.65 23.72 ± 0.85 28.80 ± 1.00 RAM (MB) 1124.24 ± 40.36 1106.92 ± 31.21 1116.28 ± 27.26 1125.34 ± 47.58 GFLOPS 89.14 ± 2.07 145.66 ± 5.17 189.91 ± 6.78 230.60 ± 8.01 Ef cientNet-B5 Size (MB): 350.67 P arameters (M): 30.00 Inferen T ime (ms) 132.28 ± 2.03 77.44 ± 0.41 60.34 ± 1.12 49.78 ± 0.44 FPS 7.56 ± 0.12 12.91 ± 0.07 16.58 ± 0.30 20.09 ± 0.18 RAM (MB) 1280.45 ± 51.11 1365.62 ± 64.42 1340.14 ± 57.63 1369.87 ± 41.35 GFLOPS 90.24 ± 1.39 154.11 ± 0.83 197.84 ± 3.56 239.75 ± 2.12 Ef cientNet-B7 Size (MB): 757.59 P arameters (M): 65.15 Inferen T ime (ms) 213.47 ± 0.95 119.06 ± 0.41 90.01 ± 0.51 74.56 ± 0.47 FPS 4.68 ± 0.02 8.40 ± 0.03 11.11 ± 0.06 13.41 ± 0.08 RAM (MB) 1854.54 ± 149.75 1751.95 ± 85.23 1896.44 ± 137.14 1840.15 ± 133.82 GFLOPS 91.99 ± 0.41 164.93 ± 0.56 218.16 ± 1.22 263.39 ± 1.66 The results indicate that the Ef cientNet-B0 model is enough to meet autonomous dri ving mi n i mum FPS (10 FPS which sho ws as green cells in the table) [34], achie v ed an inference time of 63.93 ms and 15.66 FPS on a single core, impro ving to 23.83 ms and 42.04 FPS with four cores. Dif ferent from the Ef cientNet-B0, the Ef cientNet-B5 and B7 models, which are lar ger and more comple x, require more computational resources, achie ving 132.28 ms and 7.56 FPS on a single core, b ut still reaching more than 10 FPS with tw o and three cores. The Ef cientNet-B3 model, which balances performance and ef cienc y , achie v es 89.86 ms and 11.13 FPS on a single core, impro ving to 34.76 ms and 28.80 FPS with four cores. Throughout all congurations, RAM usage remained lar gely stable for each model, ranging from ap- proximately 1.0 to 1.8 GB indicating that increased CPU parallelism primarily af fects computational speed rather than memory footprint. GFLOPS v alues and parameter counts reect each backbone’ s inherent com- ple xity , Ef cientNet-B0 is the most lightweight at 5.8M parameters and 264.36 GFLOPS (4 cores), while B7 is the lar gest at 65.15M parameters and 263.39 GFLOPS (4 cores). Model sizes on disk increase accordingly , from 68.95 MB (B0) up to 757.59 MB (B7), reinforcing the trade-of f between se gmentation accurac y and resource demand. T ak en together with our quantitati v e benchmarking and qualitati v e analysis, these edge-a w are de- plo yment results reinforce the adaptability and rob ustness of our proposed pipeline. The ability to scale from lightweight, real-time models for constrained de vices up to higher -capacity backbones for more demanding platforms, while maintaining strong se gmentation accurac y and ef cient use of computational resources, high- lights the pract ical v alue of combining pseudo-label renement, kno wledge distillation, and entrop y-based masking. Our comprehensi v e e v aluation, spanning performance metrics, visual quality , and hardw are ef - cienc y , positions this approach as a e xible and ef fecti v e solution for semantic se gmentation in di v erse real- w orld edge scenarios. Edg e-awar e distilled semantic se gmentation with pseudo-label r enement ... (No velio Putr a Indarto) Evaluation Warning : The document was created with Spire.PDF for Python.
536 ❒ ISSN: 2722-2586 4. CONCLUSION In this paper , we presented a no v el edge-a w are semantic se gmentation pipeline that le v erages pseudo- label renement, kno wledge distillat ion, and entrop y-based condence masking to train lightweight student models for deplo yment on resource-constrained de vices. Our approach systematically addresses the challenges of noisy supervision and model capacity , enabli ng ef fecti v e se gmentation in real-w orld scenarios with limited annotations. W e demonstrated that our pipeli n e can achie v e strong se gmentation performance across v ari- ous Ef cientNet backbones, with signicant impro v ements in mIoU and F1 scores through the use of rened pseudo-labels and kno wledge distillation. The entrop y-based condence masking further enhances model ro- b ustness by ltering out uncertain labels, particularly beneting smaller models that are more susceptible to noise. W e also conducted e xtensi v e edge-a w are e v aluations, sho wing that our pipeline can deli v er real-time or near real-time inference on CPUs with limited core counts, making it suitable for deplo yment in autonomous dri ving and other edge computing applications. The results indicate that e v en lightweight models can achie v e competiti v e se gmentation accurac y while maintaining lo w computational o v erhead and memory requirements, thus enabling practical deplo yment on embedded and IoT de vices. Ov erall, our ndings highlight the importance of harmonizing model capacity with supervision qual- ity and deplo yment constraints, pa ving the w ay for future research on adapti v e label renement, adv anced kno wledge transfer techniques, and conte xt-a w are deplo yment strate gies in semantic se gmentation for edge computing en vironments. By addressing the challenges of noisy labels and resource limitations, our pipeline contrib utes to the de v elopment of ef cient and ef fecti v e se gmentation solutions that can be deplo yed in real- w orld applications, particularly in autonomous dri ving scenarios where safety and reliability are paramount. FUNDING INFORMA TION This w ork is supported by Uni v ersitas Gadjah Mada under the funding of the Final Project Recognition Program (R T A, contract number: 5286/UN1.P1/PT .01.03/2024) for the Y ear 2024. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the C ontrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu No v elio Putra Indarto ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Oskar Natan ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Andi Dharma w an ✓ ✓ ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualizati on M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administrat ion V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY The data that support the ndings of this study are a v ailable on request from the corr esponding author , ON. The data, which contain information that could compromise the pri v ac y of research participants, are not publicly a v ailable due to certain restrictions. REFERENCES [1] O. Natan and J. Miura, “T o w ards compact autonomous dri ving perception with balanced learning and multi-sensor fusion, ” IEEE T r ansactions on Intellig ent T r ansportation Systems , v ol. 23, no. 9, pp. 16249–16266, Sep. 2022, doi: 10.1109/TITS.2022.3149370. IAES Int J Rob & Autom, V ol. 14, No. 4, December 2025: 528-538 Evaluation Warning : The document was created with Spire.PDF for Python.
IAES Int J Rob & Autom ISSN: 2722-2586 ❒ 537 [2] K. Ishihara, A. Kanervisto, J. Miura, and V . Hautam ¨ aki, “Multi-task learning with attention for end-to-end autonomous dri ving, ” in IEEE Confer ence on Computer V ision and P attern Reco gnition W orkshops , Jun. 2021, pp. 2896–2905, doi: 10.1109/CVPR W53098.2021.00325. [3] O. Natan and J. M iura, “Semantic se gmentation and depth estimation with RGB and D VS sensor fusion for multi-vie w dri ving perception, ” in Asian Confer ence on P attern Reco gnition , No v . 2022, pp. 352–365, doi: 10.1007/978-3-031-02375-0 26. [4] H. K. Chiu, E. Adeli, and J. C. Niebles, “Se gmenting the future, ” IEEE Robotics and A utomation Letter s , v ol. 5, no. 3, pp. 4202–4209, Jul. 2020, doi: 10.1109/LRA.2020.2992184. [5] O. Natan and J. Miura, “End-to-end autonom ous dri ving with semantic depth cloud mapping and multi-agent, ” IEEE T r ansactions on Intellig ent V ehicles , v ol. 8, no. 1, pp. 557–571, Jan. 2023, doi: 10.1109/TIV .2022.3185303. [6] L.-C. Chen, Y . Zhu, G. P apandreou, F . Schrof f, and H. Adam, “Encoder -decoder with atrous separable con v olution for semantic image se gmentation, ” in Lectur e Notes in Computer Science , v ol. 11211, pp. 833–851, 2018, doi: 10.1007/978-3-030-01234-2 49. [7] J. W ang et al. , “Deep high-resolution representation learning for visual recognition, ” IEEE T r ansactions on P attern Analysis and Mac hine Intellig ence , v ol. 43, no. 10, pp. 3349–3364, 2021, doi: 10.1109/TP AMI.2020.2983686. [8] E. Xie, W . W ang, Z. Y u, A. Anandkumar , J. M. ´ Alv arez, and P . Luo, “Se gF ormer: Simple and ef cient design for semantic se gmentation with transformers, ” Advances in Neur al Information Pr ocessing Systems , v ol. 15, pp. 12077–12090, 2021. [9] H. Bolhasani and S. J. Jassbi, “Deep l earning accelerators: A case study with MAESTR O, ” J ournal of Big Data , v ol. 7, no. 1, p. 100, 2020, doi: 10.1186/s40537-020-00377-8. [10] M. Xu, Y . Gu, S. Poslad, and S. Xue, “Optimized computation combining classication and detection netw orks with distillation, ” in International J oint Confer ence on Neur al Networks , 2021, pp. 1–8, doi: 10.1109/IJCNN52387.2021.9534331. [11] Z. Y ang, “Syner gies and challenges in the inte gration of cloud computing and deep learning: Current status, interconnectedness, and future directions, ” Highlights in Science , Engineering and T ec hnolo gy , v ol. 97, pp. 100–105, 2024, doi: 10.54097/70266446. [12] Z. Liu, M. Sun, T . Zhou, G. Huang, and T . Darrell, “Rethinking the v alue of netw ork pruning, ” International Confer ence on Learning Repr esentations , 2019. [13] B. Jacob et al. , “Quantization and training of neural netw orks for ef cient inte ger -arithmetic-only inference, ” in IEEE Confer ence on Computer V ision and P attern Reco gnition , 2018, pp. 2704–2713, doi: 10.1109/CVPR.2018.00286. [14] A. P aszk e, A. Chaurasia, S. Kim, and E. Culurciello, “ENet: A deep neural netw ork architecture for real-time semantic se gmenta- tion, ” arXiv:1606.02147 , 2016. [15] A. G. Ho w ard et al. , “MobileNets: Ef cient con v olutional neural netw orks for mobile vision applications, ” arXiv:1704.04861 , 2017. [16] M. T an and Q. V . Le, “Ef cientNet: Rethinking model scaling for con v olutional neural netw orks, ” in International Confer ence on Mac hine Learning , 2019, pp. 10691–10700. [17] G. Hint on, O. V in yals, and J. Dean, “Distilling the kno wledge in a neural netw ork, ” arXiv:1503.02531 , 2015. [18] J. Gou, B. Y u, S. J. Maybank, and D. T ao, “Kno wledge distillation: A surv e y , ” International J ournal of Computer V ision , v ol. 129, no. 6, pp. 1789–1819, 2021. [19] X. Liu, M. Neuyen, and W . Q. Y an, “V ehicle-related scene understanding using deep learning, ” in Communications in Computer and Information Science , v ol. 1180, pp. 61–73, 2020, doi: 10.1007/978-981-15-3651-9 7. [20] Y . Chen, W . Li, and L. V an Gool, “R O AD: Reality oriented adaptation for semantic se gm entation of urban scenes, ” in IEEE Con- fer ence on Computer V ision and P attern Reco gnition , 2018, pp. 7892–7901, doi: 10.1109/CVPR.2018.00823. [21] E. Arazo, D. Orte go, P . Albert, N. E. O’Connor , and K. McGuinness, “Pseudo-labeling and conrmation bias in deep semi-supervised learning, ” in International J oint Confer ence on Neur al Networks , 2020, pp. 1–8, doi: 10.1109/IJCNN48605.2020.9207304. [22] M. N. Rizv e, K. Duarte, Y . S. Ra w at, and M. Shah, “In defense of pseudo-labeling: An uncertainty-a w are pseudo-label selection frame w ork for semi-supervised learning, ” International Confer ence on Learning Repr esentations , 2021. [23] L. Lu, M. Y in, L. Fu, and F . Y ang, “Uncertainty-a w are pseudo-label and consistenc y for semi-supervised medic al image se gmenta- tion, ” Biomedical Signal Pr ocessing and Contr ol , v ol. 79, p. 104203, 2023, doi: 10.1016/j.bspc.2022.104203. [24] A. Arnab et al. , “Conditional random elds meet deep neural netw orks for semantic se gmentation: Combining probabilistic graph- ical models with deep learning for structured prediction, ” IEEE Signal Pr ocessing Ma gazine , v ol. 35, no. 1, pp. 37–52, 2018, doi: 10.1109/MSP .2017.2762355. [25] Y . W ang, Y . Huang, Q. W ang, C. Zhao, Z. Zhang, and J. Chen, “Graph-based self-training for semi-supervised deep similarity learning, ” Sensor s , v ol. 23, no. 8, p. 3944, 2023, doi: 10.3390/s23083944. [26] T . Che ng, K. Lu, H. W ang, Y . Zhao, C. Jia, and J. Xue, “Semi-supervised medical image se gmentation based on cross-pseudo- supervised guidance with high-condence pseudo labeling, ” in Medical Ima ging 2025: Ima g e Pr ocessing , v ol. 13406, SP IE, 2025, pp. 546–553, doi: 10.1117/12.3047161. [27] Q. Jiang et al. , “W eakly-supervised image semantic se gmentation based on superpix el re gion mer ging, ” Big Data and Co gnitive Computing , v ol. 3, no. 2, p. 31, 2019, doi: 10.3390/bdcc3020031. [28] M. A. M. Elhassan et al. , “Real-time semantic se gmentation for autonomous dri ving: A re vie w of CNNs, T ransformers, and be yond, ” J ournal of King Saud Univer sity – Computer and Information Sciences , v ol. 36, no. 10, p. 102226, 2024, doi: 10.1016/j.jksuci.2024.102226. [29] O. Natan and J. Miura, “DeepIPC: Deeply inte grated perception and control for an autonomous v ehicle in real en vironments, ” IEEE Access , v ol. 12, pp. 49590–49601, Apr . 2024, doi: 10.1109/A CCESS.2024.3385122. [30] M. Cordts et al. , “The Cityscapes dataset for semantic urban scene understanding, ” in IEEE Confer ence on Computer V ision and P attern Reco gnition , Jun. 2016, pp. 3213–3223, doi: 10.1109/CVPR.2016.350. [31] S. Matsuzaki, J. Miura, and H. Masuza w a, “Multi-source pseudo-label learning of semantic se gmentation for the scene recognition of agricultural mobile robots, ” Advanced Robotics , v ol. 36, no. 19, pp. 1011–1029, 2022, doi: 10.1080/01691864.2022.2109427. [32] M. J. Meni, R. T . White, M. L. M ayo, and K. R. Pil kie wicz, “Entrop y-based guidance of deep neural netw orks for accelerated Edg e-awar e distilled semantic se gmentation with pseudo-label r enement ... (No velio Putr a Indarto) Evaluation Warning : The document was created with Spire.PDF for Python.