Pith. sign in

REVIEW 4 major objections 5 minor 31 references

WS$^2$: Weakly Supervised Segmentation using Before-After Supervision in Waste Sorting

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a semantic segmentation model for unwanted objects can be trained without pixel annotations, using only video frames captured before and after a human operator removes items from a waste-sorting belt.

desk verdict Useful dataset and a plausible weak-supervision idea, but the background-removal step has a real blind spot for low-saturation objects and the evaluation is a single split with no error bars. read the letter →

arxiv 2509.06485 v1 pith:LG3SSBZV submitted 2025-09-08 cs.CV

Andrea Marelli , Alberto Foresti , Leonardo Pesce , Giacomo Boracchi , Mario Grosso This is my paper
classification cs.CV
keywords weaklysupervisedsemanticsegmentationbefore-aftersupervisionwastesortingsaliencymapsclassactivationvideoindustrialqualitycontrolSAMrefinement
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the human operator's removal action is itself a supervision signal: by comparing images of a waste-sorting belt captured before and after the operator takes items away, a network can learn to segment those unwanted objects without any pixel-level annotation. The mechanism is to train an auxiliary classifier to distinguish before from after frames; its saliency maps for the before class highlight exactly the removed objects. The paper introduces WS², the first multi-view before/after video dataset from a real sorting plant (more than 11,000 frames), and a two-stage pipeline that refines those saliency maps with SAM2 and trains SegFormer on the resulting pseudo-masks. It also shows that modeling the background as a third class suppresses background bias, and that temporal-coherent CAM methods outperform image-only ones. If the claim holds, sorting-line recognition models can be trained from operational video alone.

What carries the argument

Let "Before-After supervision" denote an auxiliary classifier trained to distinguish images taken before the operator removes items from images taken after; the saliency maps of the before class localize the removed (unwanted) objects, because those objects appear only in before frames. The two further load-bearing mechanisms are: a background-removal step that adds a third "background" class to the classifier training set, and a refinement step in which SAM2 instance masks are matched against the coarse saliency masks to produce pseudo-labels for training SegFormer.

What would settle it

Check the saliency maps on images containing only belt and no objects: if the before-class saliency still fires on the belt, background bias persists. A direct test is to train the auxiliary classifier on background-only images derived from before and after frames; if it can distinguish them above chance, the background removal has not removed the confound, and the claimed localization of unwanted items is not reliable.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that Before-After supervision—training on image pairs separated by the operator's intervention—yields a usable segmentation model for unwanted waste items. The paper demonstrates this by training a classifier to separate before from after images, deriving saliency maps for the before class, refining them through SAM2 instance masks, and using the refined maps as pseudo-labels to train a fully supervised segmentation network. A key component is a three-class training strategy in which the background is masked out of both before and after images and added as an explicit third class, preventing the classifier from latching onto near-identical b

Load-bearing premise

The background-removal step assumes that within each class the conveyor background is nearly identical and mostly gray with white reflexes, so a pixel-wise median image plus a color-saturation heuristic cleanly separates foreground objects from the belt; if the before and after cameras yield systematically different backgrounds that survive this heuristic, the auxiliary classifier can learn background cues instead of the removed objects.

Editorial extensions

If this is right

  • A segmentation model can be trained per sorting line using only recorded before/after video, removing the main labeling bottleneck for industrial sorting automation.
  • Modeling the background as a third class is claimed to be essential: without it, the classifier uses background cues and the saliency maps fail to localize unwanted items.
  • Temporal coherence in weakly supervised video segmentation is claimed to be decisive: POF-CAM, the only tested method using optical flow across frames, clearly outperforms image-only methods.
  • SAM refinement consistently improves mIoU across all methods, indicating that foundation-model instance masks can repair coarse saliency errors well enough to serve as pseudo-labels.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the supervision is defined by a change in the scene rather than by the human action itself, the same before/after paradigm could transfer to other tasks where an operator modifies an environment—assembly, inspection, food processing—not just waste sorting.
  • The background-removal heuristic (median image plus color-saturation threshold) is the component most tied to this particular belt setup; a learned background model from short belt-only video could generalize the pipeline to differently colored or patterned belts.
  • The dataset's video structure invites methods that exploit longer-range temporal context; the current POF-CAM result is likely a lower bound for what temporal-coherent weak supervision can achieve on WS².
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces WS^2, a dataset of 11,060 before/after video frames from a real plastic sorting line, and a weakly supervised segmentation pipeline. The pipeline trains an auxiliary before/after classifier using a three-class background-removal (BR) training strategy, computes saliency maps, refines them with SAM2, and uses the refined masks as pseudo-labels to train a SegFormer model. The authors benchmark several CAM-based methods (GradCAM, GradCAM++, LayerCAM, PuzzleCAM, POF-CAM, WeakTr) and report mIoU at coarse, refined, and final SegFormer stages. The best result is POF-CAM with 42.58 mIoU on before images after SegFormer training. The central claim is that a segmentation model for unwanted objects can be trained without any manual pixel annotations, using only the visual difference between before and after operator intervention.

Significance. If the central claim holds, this is a practically valuable result: sorting-line segmentation models could be trained from operational video alone, avoiding per-pixel annotation. The WS^2 dataset is a real-world, publicly released contribution with human-annotated test masks that are independent of the pseudo-labeling pipeline, which is an important validation design choice. The paper also clearly formalizes before-after supervision and demonstrates a sensible use of SAM2 refinement. However, the strength of the claim depends on the background-removal step correctly retaining all unwanted objects, and on the experimental evaluation being statistically reliable. Both of these points need additional evidence before the benchmark and the central claim can be fully accepted.

major comments (4)
  1. [Section 4 / Section 5.2] The background-removal (BR) procedure assumes foreground objects have high color saturation, while the conveyor background is gray with white reflexes. The dataset explicitly includes transparent, bluish, and opaque PET variants; transparent or white/gray unwanted objects are plausibly low-saturation and would be assigned to the background class. If so, they would never appear in the 'before (w/o bg)' training class, and the auxiliary classifier could not learn to localize them. The paper provides no quantitative evaluation of BR mask recall against the test ground truth, and no per-material or per-color breakdown of segmentation accuracy. Please add: (a) recall of M_bg on the annotated test set, stratified by material/color; (b) per-material mIoU for the final SegFormer masks; (c) sensitivity of the pipeline to the saturation/median-deviation thresholds. The detailed BR procedure is def
  2. [Section 6, Table 2] All results come from a single 80/20 train/validation split with no standard deviations, confidence intervals, or significance tests. The text states that POF-CAM 'significantly outperforms' other methods, but no statistical evidence supports this wording. Given the relatively small differences among some variants (e.g., GradCAM vs. LayerCAM at the S(TsB) stage), the ranking may not be stable. Please report results over multiple seeds or an appropriate paired significance test, and state the number of runs. This is load-bearing for the benchmark claim that temporal consistency methods outperform others.
  3. [Section 6, WeakTr paragraph] The evaluation modifies WeakTr by replacing its online refinement with the SAM-based refinement. This means the rows labeled 'WeakTr' in Table 2 do not evaluate the published WeakTr method. The modification may remove a component that contributes to WeakTr's performance, and the comparison is therefore not a fair evaluation of the state of the art. Either report the original WeakTr with its own refinement, or clearly relabel the variant and discuss the effect of the replacement on the conclusions.
  4. [Section 3 / Section 4] Before and after images are acquired by two different cameras at different positions and under 'slightly different light conditions.' Even after background removal, systematic photometric or viewpoint differences between the two cameras can provide cues to the auxiliary classifier that are unrelated to removed objects. Since the test set is drawn from the same cameras, the evaluation does not isolate the before/after removal signal from camera identity. Please provide an analysis of whether saliency maps respond to camera-specific artifacts, or report a cross-camera experiment (e.g., train on one camera pair and test on another) to demonstrate generalization.
minor comments (5)
  1. [Abstract / Page 1] Typo: 'challange' should be 'challenge'.
  2. [Figure 6 caption] The caption lists '(c)' twice; the subfigure labels should be corrected.
  3. [Section 3, Figure 4] Typo in 'we refine the the coarse masks'.
  4. [Section 6, hyperparameters] The saliency-map threshold used to produce coarse masks C(Ts) is not reported. Since the threshold directly controls the pseudo-mask quality, it should be stated and, ideally, a sensitivity analysis should be included.
  5. [Section 5.2 / Table 1] The text says 'around 9600 images' for training and 'other 1500' for test, which approximately matches the table but could be made exact (9563 and 1497) to avoid apparent inconsistency.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; before-after supervision is validated on independent human-annotated test labels, with only a minor non-load-bearing self-citation.

full rationale

The central claim—that segmentation of unwanted objects can be trained from before/after image differences—is not circular. The pipeline trains an auxiliary classifier to distinguish before and after frames, extracts saliency maps for the before class, refines them with SAM, and uses the result as pseudo-masks to train SegFormer. The final evaluation uses a test set with manually refined pixel-level annotations that are never used for training or validation. The background-removal (BR) step is a preprocessing heuristic intended to reduce background bias; it does not itself define which foreground objects are unwanted, since that distinction comes from the before/after difference. The only self-citation is POF-CAM [14], which is benchmarked as one of several competing methods and is not used to justify the framework; other methods such as PuzzleCAM also demonstrate the approach. The BR heuristic's assumption that low-saturation pixels are mostly background is a non-circular correctness risk, since it may cause objects like white or transparent plastics to be missed, but this does not make the derivation self-referential. The deferred details of the BR procedure in the supplementary material are a reproducibility concern, not evidence of circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a few domain-specific assumptions about the waste-sorting setup, mainly the background-removal heuristic and the interpretability of saliency maps. No new physical entities are introduced. Several hyperparameters are tuned for the benchmark but they are not fitted to the test set.

free parameters (4)
  • Saliency threshold for coarse masks = not reported
    Coarse masks C(TrB) are obtained by thresholding Smbef(TrB); the threshold value is not stated in the main text (Figure 4b).
  • Puzzle/POF-CAM loss weights alpha, beta = alpha=2, beta=6
    Hyperparameters tuned for the benchmark in Section 6; reported mIoU depends on these choices.
  • Background-mask saturation and median deviation thresholds = not reported
    BR masks M_bg are computed using a median image and saturation heuristic described in Section 4 and the supplementary; exact thresholds are not given in the main text.
  • Learning rates and optimizer settings = Adam lr=5e-4, SGD default for Puzzle/POF-CAM
    Chosen per method; different settings could change the reported mIoU.
assumptions (4)
  • domain assumption Saliency maps of the auxiliary classifier highlight the unwanted objects that distinguish before from after images.
    The whole pipeline relies on this interpretability assumption (Section 3).
  • domain assumption The conveyor belt background is predominantly gray with white reflexes, so foreground objects can be separated from background by high saturation and deviation from the pixel-wise median.
    Section 4, this underpins the background removal masks used to build the three-class training set.
  • domain assumption The operator's removal is the only systematic difference between before and after image sets; any occasional missed removals are random noise.
    Section 5.2, the authors state anomalies are random and infrequent.
  • standard math Classification, CAM computation, and segmentation networks behave as described in their source papers, including the SAM refinement procedure of [4].
    The paper does not re-derive these methods; it builds on them directly.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WS$^2$: Weakly Supervised Segmentation using Before-After Supervision in Waste Sorting." pith.science (2026). https://pith.science/paper/LG3SSBZV

@misc{pith2026250906485,
  author       = {Pith},
  title        = {Pith review of: WS$^2$: Weakly Supervised Segmentation using Before-After Supervision in Waste Sorting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LG3SSBZV}},
  note         = {Machine review of arXiv:2509.06485}
}
abstract

In industrial quality control, to visually recognize unwanted items within a moving heterogeneous stream, human operators are often still indispensable. Waste-sorting stands as a significant example, where operators on multiple conveyor belts manually remove unwanted objects to select specific materials. To automate this recognition problem, computer vision systems offer great potential in accurately identifying and segmenting unwanted items in such settings. Unfortunately, considering the multitude and the variety of sorting tasks, fully supervised approaches are not a viable option to address this challange, as they require extensive labeling efforts. Surprisingly, weakly supervised alternatives that leverage the implicit supervision naturally provided by the operator in his removal action are relatively unexplored. In this paper, we define the concept of Before-After Supervision, illustrating how to train a segmentation network by leveraging only the visual differences between images acquired \textit{before} and \textit{after} the operator. To promote research in this direction, we introduce WS$^2$ (Weakly Supervised segmentation for Waste-Sorting), the first multiview dataset consisting of more than 11 000 high-resolution video frames captured on top of a conveyor belt, including "before" and "after" images. We also present a robust end-to-end pipeline, used to benchmark several state-of-the-art weakly supervised segmentation methods on WS$^2$.

Figures

Figures reproduced from arXiv: 2509.06485 by the authors.

Figure 1
Figure 1. In a waste sorting plant, two cameras, C1 and C2, are placed along a conveyor belt where a human operator manually re￾moves unwanted objects (red) from a mixed waste stream, leaving on the belt only wanted ones (green). C1 captures the belt sec￾tion before the HO’s intervention, while C2 captures the section after, where only wanted objects remain. We aim to train, with￾out any pixel-wise annotation nor additional s… view at source ↗
Figure 2
Figure 2. (a) “Before” frame showing both wanted and unwanted waste items on the conveyor belt. (b) Annotation overlay: green outlines mark the wanted items to keep (semi-transparent PET bottles), while red outlines indicate unwanted waste to be removed (all other plastic materials). (c) “After” frame in which only the wanted items remain. The absence of the unwanted waste, present only in the “before” frames, serves as super… view at source ↗
Figure 3
Figure 3. Overview of the first stage in our pipeline. We train the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: (a) In the WS module, we leverage the trained classifier [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: (a) Before image saliency map (Sm(I B)) from a clas￾sifier (POF-CAM [14]) trained with standard before-after images with the background included (as in [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Qualitative examples of SMs obtained using different [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 23 canonical work pages

  1. [1]

    Attach dataset: Annotated two-handed assembly actions for human action understanding

    Dustin Aganian, Benedict Stephan, Markus Eisenbach, Corinna Stretz, and Horst-Michael Gross. Attach dataset: Annotated two-handed assembly actions for human action understanding. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11367–11373. IEEE, 2023. 3

    work page 2023

  2. [2]

    Zerowaste dataset: To- wards deformable object segmentation in cluttered scenes

    Dina Bashkirova, Mohamed Abdelfattah, Ziliang Zhu, James Akl, Fadi Alladkani, Ping Hu, Vitaly Ablavsky, Berk Calli, Sarah Adel Bargal, and Kate Saenko. Zerowaste dataset: To- wards deformable object segmentation in cluttered scenes. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 21147–21157, 2022. 2, 3, 4

    work page 2022

  3. [3]

    Grad-cam++: General- ized gradient-based visual explanations for deep convolu- tional networks

    Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: General- ized gradient-based visual explanations for deep convolu- tional networks. In2018 IEEE winter conference on appli- cations of computer vision (WACV), pages 839–847. IEEE,

    work page

  4. [4]

    Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation.arXiv preprint arXiv:2305.05803, 2023

    Tianle Chen, Zheda Mai, Ruiwen Li, and Wei-lun Chao. Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation.arXiv preprint arXiv:2305.05803, 2023. 5, 7

    arXiv 2023

  5. [5]

    Timbersleuth: Visual anomaly detec- tion with human feedback for mitigating the illegal timber trade.Information Visualization, 22(3):223–245, 2023

    Debanjan Datta, Nathan Self, John Simeone, Amelia Mead- ows, Willow Outhwaite, Linda Walker, Niklas Elmqvist, and Naren Ramakrishnan. Timbersleuth: Visual anomaly detec- tion with human feedback for mitigating the illegal timber trade.Information Visualization, 22(3):223–245, 2023. 3

    work page 2023

  6. [6]

    Trash- icra19: A bounding box labeled dataset of underwater trash

    Michael S Fulton, Jungseok Hong, and Junaed Sattar. Trash- icra19: A bounding box labeled dataset of underwater trash

    work page

  7. [7]

    Trash- can: A semantically-segmented dataset towards visual de- tection of marine debris.arXiv preprint arXiv:2007.08097,

    Jungseok Hong, Michael Fulton, and Junaed Sattar. Trash- can: A semantically-segmented dataset towards visual de- tection of marine debris.arXiv preprint arXiv:2007.08097,

    arXiv 2007

  8. [8]

    Egotaskqa: Understanding human tasks in egocentric videos

    Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. Advances in Neural Information Processing Systems, 35: 3343–3360, 2022. 3

    work page 2022

Show all 31 references
  1. [9]

    Layercam: Exploring hierarchical class activation maps for localization.IEEE Transactions on Image Processing, 30:5875–5888, 2021

    Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization.IEEE Transactions on Image Processing, 30:5875–5888, 2021. 7, 8

    2021

  2. [10]

    Puzzle-cam: Improved local- ization via matching partial and full features

    Sanghyun Jo and In-Jae Yu. Puzzle-cam: Improved local- ization via matching partial and full features. In2021 IEEE international conference on image processing (ICIP), pages 639–643. IEEE, 2021. 7, 8

    2021

  3. [11]

    Maria Koskinopoulou, Fredy Raptopoulos, George Pa- padopoulos, Nikitas Mavrakis, and Michail Maniadakis. Robotic waste sorting technology: Toward a vision-based categorization system for the industrial robotic separation of recyclable waste.IEEE Robotics & Automation Magazine,...

    2021

  4. [12]

    Trashbox: trash detection and classification using quantum transfer learning

    Nikhil Venkat Kumsetty, Amith Bhat Nekkare, Sowmya Ka- math, et al. Trashbox: trash detection and classification using quantum transfer learning. In2022 31st Conference of Open Innovations Association (FRUCT), pages 125–130. IEEE, 2022. 3, 4

    2022

  5. [13]

    Learning to recognize procedural activities with distant supervision

    Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853–13863, 2022. 3

    2022

  6. [14]

    Temporal-consistent cams for weakly su- pervised video segmentation in waste sorting

    Andrea Marelli, Luca Magri, Federica Arrigoni, and Gia- como Boracchi. Temporal-consistent cams for weakly su- pervised video segmentation in waste sorting. InEuropean Conference on Computer Vision. Springer, 2025. 2, 4, 6, 7, 8

    2025

  7. [15]

    The brio-ta dataset: Understanding anomalous assembly process in manufacturing

    Kosuke Moriwaki, Gaku Nakano, and Tetsuo Inoshita. The brio-ta dataset: Understanding anomalous assembly process in manufacturing. In2022 IEEE International Conference on Image Processing (ICIP), pages 1991–1995. IEEE, 2022. 3

    1991

  8. [16]

    Aquavision: Automating the detection of waste in water bodies using deep transfer learn- ing.Case Studies in Chemical and Environmental Engineer- ing, 2:100026, 2020

    Harsh Panwar, PK Gupta, Mohammad Khubeb Siddiqui, Ruben Morales-Menendez, Prakhar Bhardwaj, Sudhansh Sharma, and Iqbal H Sarker. Aquavision: Automating the detection of waste in water bodies using deep transfer learn- ing.Case Studies in Chemical and Environmental Engineer- in...

    2020

  9. [17]

    Taco: Trash annotations in context for litter detection

    PF Proenc ¸a and P Sim ˜oes. Taco: Trash annotations in context for litter detection. arxiv 2020.arXiv preprint arXiv:2003.06975, 2003. 3, 4

    2020 arXiv

  10. [18]

    The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain

    Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer V...

    2021

  11. [19]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024. 5, 7

    2024 arXiv

  12. [20]

    Industreal: A dataset for procedure step recognition handling execution errors in egocentric videos in an industrial-like setting

    Tim J Schoonbeek, Tim Houben, Hans Onvlee, Fons Van der Sommen, et al. Industreal: A dataset for procedure step recognition handling execution errors in egocentric videos in an industrial-like setting. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer...

    2024

  13. [21]

    Grad-cam: Visual explanations from deep networks via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. InProceedings of the IEEE in- ternational conference on computer vision, pages 618–626,

  14. [22]

    Automation of waste sorting with deep learning

    Joao Sousa, Ana Rebelo, and Jaime S Cardoso. Automation of waste sorting with deep learning. In2019 XV Workshop de visao Computacional (WVC), pages 43–48. IEEE, 2019. 3, 4

    2019

  15. [23]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through at- tention. InInternational conference on machine learning, pages 10347–10357. PMLR, 2021. 7

    2021

  16. [24]

    triangle of death

    Maria Triassi, Rossella Alfano, Maddalena Illario, Antonio Nardone, Oreste Caporale, and Paolo Montuori. Environ- mental pollution from illegal waste disposal and health ef- fects: A review on the “triangle of death”.International Journal of Environmental Research and Public H...

    2015

  17. [25]

    Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in neural information processing systems, 34: 12077–12090, 2021

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in neural information processing systems, 34: 12077–12090, 2021. 5, 8

    2021

  18. [26]

    Egocentric video task translation

    Zihui Xue, Yale Song, Kristen Grauman, and Lorenzo Tor- resani. Egocentric video task translation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2310–2320, 2023. 3

    2023

  19. [27]

    Classification of trash for re- cyclability status.CS229 project report, 2016(1):3, 2016

    Mindy Yang and Gary Thung. Classification of trash for re- cyclability status.CS229 project report, 2016(1):3, 2016. 3, 4

    2016

  20. [28]

    Hierarchical waste detection with weakly supervised segmentation in images from recycling plants.Engineering Applications of Artificial Intelligence, 128:107542, 2024

    Dmitry Yudin, Nikita Zakharenko, Artem Smetanin, Roman Filonov, Margarita Kichik, Vladislav Kuznetsov, Dmitry Larichev, Evgeny Gudov, Semen Budennyy, and Aleksandr Panov. Hierarchical waste detection with weakly supervised segmentation in images from recycling plants.Engineeri...

    2024

  21. [29]

    Learning deep features for discrimina- tive localization

    Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discrimina- tive localization. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2921–2929,

  22. [30]

    Weaktr: Exploring plain vision transformer for weakly-supervised semantic segmen- tation.arXiv preprint arXiv:2304.01184, 2023

    Lianghui Zhu, Yingyue Li, Jiemin Fang, Yan Liu, Hao Xin, Wenyu Liu, and Xinggang Wang. Weaktr: Exploring plain vision transformer for weakly-supervised semantic segmen- tation.arXiv preprint arXiv:2304.01184, 2023. 7, 8

    2023

  23. [31]

    Cross- task weakly supervised learning from instructional videos

    Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross- task weakly supervised learning from instructional videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3537–3545, 2019. 3

    2019

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.