Pith. sign in

REVIEW 3 major objections 6 minor 50 references

Depth-aware tiny networks for affordance segmentation sit on the Pareto front of accuracy versus size and run real-time on phone-scale power.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 12:02 UTC pith:JZJRGA2P

load-bearing objection Useful mid-tier RGB-D AS engineering with a real search-space tweak and honest Jetson numbers; Pareto claim holds on their cropped protocol, transfer to full wearable scenes is the open caveat. the 3 major comments →

arxiv 2607.28293 v1 pith:JZJRGA2P submitted 2026-07-30 cs.CV cs.LGcs.RO

Filling the Pareto-Optimal Front for Affordance Segmentation on Embedded Devices Using RGB-D Cameras

Edoardo Ragusa , Giovanni Paolo Canuti , Simone Lugani , Rodolfo Zunino , Paolo Gastaldo This is my paper
classification cs.CV cs.LGcs.RO
keywords affordance segmentationRGB-Dhardware-aware NASembedded systemstiny CNNswearable robotsJetson NanoPareto front
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Wearable robots need to segment objects into graspable parts from on-board cameras, but tiny networks that fit microcontrollers often fail in hard lighting and framing, while large models are too heavy. This paper shows that portable RGB-D cameras can close the gap if depth is integrated carefully. It offers two practical routes: a hardware-aware neural architecture search whose search space builds separate RGB and depth branches that merge early and keep multi-resolution skip links, and a cheap fine-tuning trick that adds a single 4-to-3 convolutional adapter so any pre-trained RGB network can ingest depth. On two standard affordance datasets the resulting models dominate or sit on the Pareto front of pixel accuracy versus parameters and FLOPs relative to prior tiny and mid-range baselines, and a Jetson Nano plus RealSense prototype runs them in real time under a smartphone-battery energy budget.

Core claim

A reformulated hardware-aware NAS with an RGB-D multi-branch search space, together with a one-layer fine-tuning adapter that maps four-channel RGB-D into three-channel RGB backbones, produce affordance-segmentation networks that, in most cases, occupy the Pareto-optimal front balancing generalization against model size and FLOPs, outperforming existing tiny and hardware-efficient baselines while sustaining real-time inference on a Jetson Nano within smartphone-battery power limits.

What carries the argument

The RGB-D search space: independent MobileNetV3-style encoder stacks for RGB and depth that fuse early into a shared backbone, plus an aggregator that freely wires multi-resolution skip connections from any encoder stage; the complementary 4→3 convolutional adapter that lets any frozen RGB network ingest depth with almost no extra cost.

Load-bearing premise

That accuracy measured on tightly cropped foreground object patches is a fair stand-in for real wearable use, where the user must still frame the object and background, lighting, and localization errors remain.

What would settle it

Deploy the same networks end-to-end on a wearable arm with live RealSense framing (no oracle crops) and measure whether grasp/don’t-grasp pixel accuracy and Pareto ranking versus the tiny baselines collapse under realistic mis-framing, clutter, and illumination.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Mid-range embedded accelerators become a practical home for affordance segmentation once depth is fused this way, rather than forcing a choice between microcontroller-tiny models and GPU-scale ones.
  • Existing RGB-only affordance networks can be upgraded to RGB-D with a single learned adapter and negligible extra FLOPs or parameters.
  • The same multi-branch search space can be reused to trade accuracy against power when input resolution is lowered, keeping average accuracy above 0.9 even near microcontroller FLOP budgets.
  • A Jetson-class board plus RealSense already meets real-time and smartphone-battery constraints for on-device affordance maps that can feed semi-autonomous grasp control.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the NAS consistently prefers early fusion and large kernels with stride-1, geometry-preserving low-level features appear more valuable than deep separate streams for this task.
  • The same adapter-plus-fine-tune pattern should transfer to other dense prediction heads (keypoints, grasp quality) that already have strong RGB checkpoints.
  • Once object localization is added as promised in the conclusion, the Pareto models could become drop-in perception modules for closed-loop prosthetic or exoskeleton controllers without cloud offload.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper presents two design strategies for affordance segmentation (AS) from RGB-D on embedded hardware for wearable robots: (i) a hardware-aware NAS with a new multi-branch search space (independent RGB and D encoder stacks, early-merge backbone, aggregator with searchable multi-resolution skips), and (ii) a lightweight fine-tuning path that inserts a learned 4→3 convolutional adapter so pre-trained RGB backbones (MobileNetV3) can consume RGB-D. Both target mid-range accelerators rather than pure tinyML or unconstrained GPUs. On UMD and IIT (collapsed to 3-class pixel-wise grasp / don’t-grasp / background), the resulting models often lie on the Pareto front of accuracy versus parameters/FLOPs relative to tiny baselines, CustomMobV3, and related efficient nets; a Jetson Nano + RealSense prototype demonstrates real-time inference and smartphone-battery-compatible power after TRT/FP16 optimization.

Significance. Depth for portable AS and mid-range (accelerator-class) architectures are genuinely underexplored relative to unconstrained and microcontroller extremes. The multi-branch RGB-D search space, the near-zero-overhead fine-tuning adapter, explicit Pareto scatter plots (Figs. 5–6), resolution ablation, per-object depth-gain analysis, and hardware-in-the-loop latency/memory/power numbers on a concrete prototype are concrete, reusable contributions. If the reported trade-offs hold under more realistic framing, the work supplies practical design recipes and baselines for semi-autonomous grasping pipelines on wearable devices.

major comments (3)
  1. [§IV-A, §V, Figs. 5–6] §IV-A and §V: All accuracy and Pareto results (Figs. 2–8, Tables II–IV) are obtained exclusively on tightly cropped foreground squares extracted via ground-truth masks; framing and background are left to the user and deferred to future work. The central claim—Pareto-optimal RGB-D AS for wearable near-sensor control—therefore rests on a distribution that omits background clutter, localization error, and mis-framing that real control pipelines must handle. Device latency/power remain valid, but it is unclear whether depth gains (Figs. 2, 8) or dominance over tiny baselines (Figs. 5–6) survive once full frames are used. Either add full-frame / localization-aware experiments, or substantially qualify the wearable/Pareto claims and title/abstract so they match the evaluated distribution.
  2. [§IV-D, Fig. 4, Table IV] §IV-D / Fig. 4 (IIT): On the more varied IIT benchmark, depth yields little or no average gain for the NAS models, and fine-tuned Mob3-RGB-D is consistently stronger than NAS-RGB-D; large ImageNet-pretrained models dominate. The paper attributes this to feature robustness and parameter count but does not quantify whether the RGB-D search space under-explores capacity under domain shift, whether early merge (Table IV) discards useful depth cues, or whether the FLOPs threshold (MobileNetV3 @ 128²) simply starves the dual-branch design. A short ablation (later merge, higher FLOPs budget, or frozen vs. joint fine-tuning of the adapter) is needed to support the claim that the proposed methods “in most cases” fill the Pareto front across realistic conditions.
  3. [§III-A, Eq. (1)] §III-A, Eq. (1): The evolutionary search is described at a high level (100 generations, children via random mutations, selection by a “predefined metric,” early-stop child training for 10 epochs). The selection metric (validation loss only? accuracy–FLOPs scalarization?), mutation operators, population size, and whether φ(a) is a proxy or measured latency are not specified. Without these, the HW-NAS contribution is hard to reproduce or to distinguish from a lucky manual design. Please state the exact fitness function, mutation set, and any hardware-in-the-loop measurements used during search.
minor comments (6)
  1. [Fig. 1] Fig. 1 caption is partly in Italian (“per la segmentazione di affordance near-sensor usando camere RGB-D”); replace with English for consistency.
  2. [Table IV] Table IV lists two rows both labeled “b5” with different activations; clarify numbering of the backbone blocks.
  3. [§IV-D, Figs. 3–4] Figs. 3–4 omit TinySeg bars because accuracy is very low; state the numeric scores in the caption or text so the comparison remains complete.
  4. [§IV-A, Table II, Fig. 7] FLOPs units are written inconsistently (KFLOPS threshold vs. MFLOPS/M in tables and Fig. 7); standardize and confirm the MobileNetV3 reference value (528 KFLOPS appears to be a typographical slip for ~528 MFLOPS).
  5. [§IV-G] §IV-G power estimates assume continuous inference; a short note on duty-cycled use (user-triggered AS) would better support the “smartphone battery” claim.
  6. [Figs. 5, Tables III–IV] Minor typos: “Avera e” in Fig. 5 axis; “grb1” in Table IV; “hd sig./hd sgm” activation abbreviations should be defined once.

Circularity Check

0 steps flagged

Empirical systems paper: NAS/fine-tuning results and Pareto claims are measured on held-out data and hardware, not derived by construction from inputs or self-citations.

full rationale

The paper’s load-bearing claims are experimental outcomes, not first-principles derivations. HW-NAS (Eq. 1, new multi-branch RGB-D search space in §III-A, Tables III–IV) and the 4→3 conv fine-tuning adapter (§III-B) produce concrete architectures that are trained and scored on held-out UMD/IIT test splits and then profiled for FLOPs, parameters, latency, memory, and power on a Jetson Nano + RealSense prototype (§IV). Pareto dominance (Figs. 5–6) is an observational ranking of those measured points against external and prior baselines; it is not forced by the definition of the search space, the FLOPs threshold (set to MobileNetV3’s cost), or the evolutionary parent-selection rule. Self-citations ([8], [12], [14], etc.) supply motivation, the CustomMobV3/HW-tiny baselines, and the prior RGB-only NAS setup, but the new RGB-D accuracy gains, architecture structures, and device measurements are independent empirical content. No quantity is defined in terms of the target it is said to predict, no uniqueness theorem is imported to forbid alternatives, and no fitted parameter is relabeled a prediction. Circularity score is therefore 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

Central claims rest on standard DL/embedded assumptions plus author-chosen NAS budgets and a simplified three-class AS task on cropped objects. No new physical entities; the “invented” pieces are design constructs (search space, adapter layer) whose value is purely empirical.

free parameters (5)
  • NAS FLOPs threshold = stated as MobileNetV3@128² (text: 528 KFLOPS; tables ~528M scale)
    Admissible architectures constrained to MobileNetV3 Keras FLOPs at 128×128; sets the hardware-aware feasible set and thus who can appear on the reported Pareto front.
  • NAS generations and child training budget = 100 generations; 10 then 100 epochs; lr 1e-3
    Search runs 100 generations with children trained 10 epochs under early stopping; finalist retrained 100 epochs. These choices control which architecture is declared best.
  • Evolutionary mutation/selection metric
    Parent/child evolutionary NAS without a formally stated multi-objective scalarization beyond validation loss and ϕ(a)<Th; selection metric details affect claimed optimality.
  • Input resolutions tested = 128×128 default; 64 and 48 ablated
    128/64/48 used to trade accuracy vs FLOPs; resolution is a free design knob presented as the main lever toward MCU-like budgets.
  • 3-class affordance collapse = grasp / don’t grasp / background
    All grasp affordances merged to one class and others to don’t-grasp plus background, following prior wearable AS setup; defines the accuracy numbers on the Pareto plots.
axioms (5)
  • domain assumption FLOPs and parameter count are adequate device-agnostic proxies for embedded feasibility; inference-time constraint ϕ(a)<Th can be enforced mainly via FLOPs at search time.
    Stated in §III-A and §IV-A/E; deployment later measures real FPS/power but search does not use hardware-in-the-loop latency.
  • domain assumption Depth can be treated as a noisy single-channel image correlated with RGB, so CNN encoder blocks (MobileNetV3-like without SE) and early fusion are appropriate.
    §III-A–B; justifies both the dual-backbone search space and the claim that fine-tuning low-level filters is safe.
  • ad hoc to paper Cropped foreground squares from dataset masks are a sufficient training/evaluation distribution for near-sensor wearable AS.
    §IV-A processing pipeline; conclusion admits framing/localization left to the user.
  • standard math Holdout validation from the training split and never-tuned test sets yield unbiased generalization estimates for architecture comparison.
    §IV-A; standard ML practice, no nested search-bias correction discussed for NAS.
  • domain assumption Real-time for human-in-the-loop control is satisfied by ~8–16 FPS and multi-hour runtime on a ~5000 mAh phone battery at ~7 W average.
    §IV-G power and FPS discussion; application-level axiom linking measured rates to usability.
invented entities (2)
  • RGB-D multi-branch HW-NAS search space (independent RGB/D encoder stacks, merge backbone, aggregator with searchable multi-resolution skips) no independent evidence
    purpose: Let evolutionary search discover small RGB-D AS nets under FLOPs/flash-style constraints instead of hand-merging modalities.
    Core methodological object in §III-A and Tables I/III/IV; value is empirical only.
  • 4-channel to 3-channel learned preprocessing conv for depth-aware fine-tuning of RGB backbones no independent evidence
    purpose: Inject depth into pre-trained RGB AS models with negligible extra FLOPs/params.
    §III-B; standard adapter idea specialized to this pipeline, not an external physical entity.

pith-pipeline@v1.2.0-daily-grok45 · 21934 in / 3866 out tokens · 97711 ms · 2026-07-31T12:02:31.137202+00:00 · methodology

0 comments
read the original abstract

While depth sensors have the potential to complement RGB data for affordance segmentation in wearable robots, their usage seems to remain underexplored. The paper proposes two approaches: a reformulated version of hardware-aware neural architecture search, endowed with a newly designed search space to integrate depth (D) information into small-sized deep networks, and a dedicated fine-tuning approach, including a preprocessing layer to merge depth information with RGB data and make it compatible with conventional architectures. In both cases, those methods aim to generate solutions that benefit from modern (portable) hardware accelerators and overcome existing tiny-like approaches, which often fail to tackle critical scenarios due to the severe constraints set by the supporting hardware. Extensive experiments on a pair of real-world datasets demonstrate the effectiveness of the proposed method as compared with existing solutions. The approach presented in the paper generates, in most cases, solutions that identify the Pareto optimal front to balance generalization performance and hardware requirements. The paper also describes the supporting prototype, including a Jetson Nano board and a RealSense RGB-D camera. When considering the energy profile of the device, the overall system can attain real-time performances within an energy budget that is compatible with standard batteries, such as those used in smartphones.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 7 linked inside Pith

  1. [1]

    Wearable supernumerary robotic limb system using a hybrid control approach based on motor imagery and object detection,

    Z. Tang, L. Zhang, X. Chen, J. Ying, X. Wang, and H. Wang, “Wearable supernumerary robotic limb system using a hybrid control approach based on motor imagery and object detection,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 30, pp. 1298–1309, 2022

    2022

  2. [2]

    Current rates of prosthetic usage in upper-limb amputees–have innovations had an impact on device acceptance?

    S. Salminger, H. Stino, L. H. Pichler, C. Gstoettner, A. Sturma, J. A. Mayer, M. Szivak, and O. C. Aszmann, “Current rates of prosthetic usage in upper-limb amputees–have innovations had an impact on device acceptance?”Disability and Rehabilitation, vol. 44, no. 14, pp. 3708– 3713, 2022

    2022

  3. [3]

    Real-time radar-based gesture detection and recognition built in an edge-computing platform,

    Y . Sun, T. Fei, X. Li, A. Warnecke, E. Warsitz, and N. Pohl, “Real-time radar-based gesture detection and recognition built in an edge-computing platform,”IEEE Sensors Journal, vol. 20, no. 18, pp. 10 706–10 716, 2020

    2020

  4. [4]

    The future of bionic limbs: The untapped synergy of signal processing, control, and wireless connectivity,

    F. Chiariotti, P. Mamidanna, S. Suman, ˇC. Stefanovi ´c, D. Farina, P. Popovski, and S. Do ˇsen, “The future of bionic limbs: The untapped synergy of signal processing, control, and wireless connectivity,”IEEE Signal Processing Magazine, vol. 41, no. 4, pp. 58–75, 2024

    2024

  5. [5]

    A survey of teleceptive sensing for wearable assistive robotic devices,

    N. E. Krausz and L. J. Hargrove, “A survey of teleceptive sensing for wearable assistive robotic devices,”Sensors, vol. 19, no. 23, p. 5238, 2019

    2019

  6. [6]

    Explorations of autonomous prosthetic grasping via proximity vision and deep learning,

    E. Mastinu, A. Coletti, J. van den Berg, and C. Cipriani, “Explorations of autonomous prosthetic grasping via proximity vision and deep learning,” IEEE Transactions on Medical Robotics and Bionics, 2024

    2024

  7. [7]

    Continuous semi-autonomous prosthesis control using a depth sensor on the hand,

    M. N. Castro and S. Dosen, “Continuous semi-autonomous prosthesis control using a depth sensor on the hand,”Frontiers in Neurorobotics, vol. 16, 2022

    2022

  8. [8]

    Hardware-aware affordance detection for application in portable embedded systems,

    E. Ragusa, C. Gianoglio, S. Dosen, and P. Gastaldo, “Hardware-aware affordance detection for application in portable embedded systems,” IEEE Access, vol. 9, pp. 123 178–123 193, 2021

    2021

  9. [9]

    Learning affordance segmentation for real-world robotic manipulation via synthetic images,

    F.-J. Chu, R. Xu, and P. A. Vela, “Learning affordance segmentation for real-world robotic manipulation via synthetic images,”IEEE Robotics and Automation Letters, vol. 4, no. 2, pp. 1140–1147, 2019

    2019

  10. [10]

    Synergies between affordance and geometry: 6-dof grasp detection via implicit representa- tions,

    Z. Jiang, Y . Zhu, M. Svetlik, K. Fang, and Y . Zhu, “Synergies between affordance and geometry: 6-dof grasp detection via implicit representa- tions,”arXiv preprint arXiv:2104.01542, 2021

    Pith/arXiv arXiv 2021

  11. [11]

    Towards visual affordance learning: A benchmark for affordance segmentation and recognition,

    Z. O. Khalifa and S. A. A. Shah, “Towards visual affordance learning: A benchmark for affordance segmentation and recognition,”arXiv preprint arXiv:2203.14092, 2022

    Pith/arXiv arXiv 2022

  12. [12]

    Affordance seg- mentation using tiny networks for sensing systems in wearable robotic devices,

    E. Ragusa, S. Dosen, R. Zunino, and P. Gastaldo, “Affordance seg- mentation using tiny networks for sensing systems in wearable robotic devices,”IEEE Sensors Journal, 2023

    2023

  13. [13]

    A survey on rgb-d datasets,

    A. Lopes, R. Souza, and H. Pedrini, “A survey on rgb-d datasets,” Computer Vision and Image Understanding, vol. 222, p. 103489, 2022

    2022

  14. [14]

    Affordance segmentation using rgb-d sensors for application in portable embedded systems,

    E. Ragusa, M. P. Ghezzi, R. Zunino, and P. Gastaldo, “Affordance segmentation using rgb-d sensors for application in portable embedded systems,” inApplications in Electronics Pervading Industry, Environ- ment and Society: APPLEPIES 2022. Springer, 2023, pp. 109–116

    2022

  15. [15]

    Embedded deep learning accel- erators: A survey on recent advances,

    G. Akkad, A. Mansour, and E. Inaty, “Embedded deep learning accel- erators: A survey on recent advances,”IEEE Transactions on Artificial Intelligence, 2023

    2023

  16. [16]

    Machine learning for microcontroller-class hardware-a review,

    S. S. Saha, S. S. Sandha, and M. Srivastava, “Machine learning for microcontroller-class hardware-a review,”IEEE Sensors Journal, 2022

    2022

  17. [17]

    Affordance detection of tool parts from geometric features,

    A. Myers, C. L. Teo, C. Ferm ¨uller, and Y . Aloimonos, “Affordance detection of tool parts from geometric features,” in2015 IEEE Interna- tional Conference on Robotics and Automation (ICRA). IEEE, 2015, pp. 1374–1381

    2015

  18. [18]

    Detecting object affordances with convolutional neural networks,

    A. Nguyen, D. Kanoulas, D. G. Caldwell, and N. G. Tsagarakis, “Detecting object affordances with convolutional neural networks,” in 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2016, pp. 2765–2770

    2016

  19. [19]

    Resource-aware object classification and segmentation for semi-autonomous grasping with prosthetic hands,

    F. Hundhausen, D. Megerle, and T. Asfour, “Resource-aware object classification and segmentation for semi-autonomous grasping with prosthetic hands,” in2019 IEEE-RAS 19th International Conference on Humanoid Robots (Humanoids). IEEE, 2019, pp. 215–221

    2019

  20. [20]

    Semi-autonomous control of prosthetic hands based on multimodal sensing, human grasp demonstration and user intention,

    J. Starke, P. Weiner, M. Crell, and T. Asfour, “Semi-autonomous control of prosthetic hands based on multimodal sensing, human grasp demonstration and user intention,”Robotics and Autonomous Systems, vol. 154, p. 104123, 2022

    2022

  21. [21]

    Designing prosthetic hands with embodied intelligence: The kit prosthetic hands,

    P. Weiner, J. Starke, S. Rader, F. Hundhausen, and T. Asfour, “Designing prosthetic hands with embodied intelligence: The kit prosthetic hands,” Frontiers in Neurorobotics, vol. 16, 2022

    2022

  22. [22]

    Sensor fusion and computer vision for context-aware control of a multi degree-of-freedom prosthesis,

    M. Markovic, S. Dosen, D. Popovic, B. Graimann, and D. Farina, “Sensor fusion and computer vision for context-aware control of a multi degree-of-freedom prosthesis,”Journal of neural engineering, vol. 12, no. 6, p. 066022, 2015

    2015

  23. [23]

    Affordance-centric policy learning: Sample efficient and generalis- able robot policy learning using affordance-centric task frames,

    K. Rana, J. Abou-Chakra, S. Garg, R. Lee, I. Reid, and N. Suenderhauf, “Affordance-centric policy learning: Sample efficient and generalis- able robot policy learning using affordance-centric task frames,”arXiv preprint arXiv:2410.12124, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    A review of sensory feedback in upper- limb prostheses from the perspective of human motor control,

    J. W. Sensinger and S. Dosen, “A review of sensory feedback in upper- limb prostheses from the perspective of human motor control,”Frontiers in Neuroscience, vol. 14, 2020

    2020

  25. [25]

    Toward affordance detection and ranking on novel objects for real-world robotic manipulation,

    F.-J. Chu, R. Xu, L. Seguin, and P. A. Vela, “Toward affordance detection and ranking on novel objects for real-world robotic manipulation,”IEEE Robotics and Automation Letters, vol. 4, no. 4, pp. 4070–4077, 2019

    2019

  26. [26]

    An affordance keypoint detection network for robot manipulation,

    R. Xu, F.-J. Chu, C. Tang, W. Liu, and P. A. Vela, “An affordance keypoint detection network for robot manipulation,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 2870–2877, 2021

    2021

  27. [27]

    Sgsin: Simultaneous grasp and suction inference network via attention-based affordance learning,

    W. Wang, H. Zhu, and M. H. Ang Jr, “Sgsin: Simultaneous grasp and suction inference network via attention-based affordance learning,”IEEE Transactions on Industrial Electronics, 2024

    2024

  28. [28]

    Visual affordance and function understanding: A survey,

    M. Hassanin, S. Khan, and M. Tahtali, “Visual affordance and function understanding: A survey,”ACM Computing Surveys (CSUR), vol. 54, no. 3, pp. 1–35, 2021

    2021

  29. [29]

    Ganhand: Predicting human grasp affordances in multi-object scenes,

    E. Corona, A. Pumarola, G. Alenya, F. Moreno-Noguer, and G. Rogez, “Ganhand: Predicting human grasp affordances in multi-object scenes,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 5031–5041

    2020

  30. [30]

    Affordance labeling and exploration: A manifold-based approach,

    ˙I. ¨Ozc ¸il and A. B. Koku, “Affordance labeling and exploration: A manifold-based approach,”arXiv preprint arXiv:2407.15479, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    An affordance detection pipeline for resource-constrained AUTHORet al.: PREP ARA TION OF P APERS FOR IEEE TRANSACTIONS AND JOURNALS 11 devices,

    T. Apicella, A. Cavallaro, R. Berta, P. Gastaldo, F. Bellotti, and E. Ragusa, “An affordance detection pipeline for resource-constrained AUTHORet al.: PREP ARA TION OF P APERS FOR IEEE TRANSACTIONS AND JOURNALS 11 devices,” in2021 28th IEEE International Conference on Electronics, Circuits, and Systems (ICECS). IEEE, pp. 1–6

  32. [32]

    Lightweight neural networks for affordance segmentation: Enhancement of the decoder module,

    S. Lugani, E. Ragusa, R. Zunino, and P. Gastaldo, “Lightweight neural networks for affordance segmentation: Enhancement of the decoder module,” inInternational Conference on Applications in Electronics Pervading Industry, Environment and Society. Springer, 2023, pp. 437– 443

    2023

  33. [33]

    3d affordancenet: A benchmark for visual object affordance understanding,

    S. Deng, X. Xu, C. Wu, K. Chen, and K. Jia, “3d affordancenet: A benchmark for visual object affordance understanding,” inproceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 1778–1787

    2021

  34. [34]

    Are dense labels always necessary for 3d object detection from point cloud?

    C. Gao, C. Liu, J. Shu, F. Liu, J. Liu, L. Yang, X. Gao, and D. Meng, “Are dense labels always necessary for 3d object detection from point cloud?”arXiv preprint arXiv:2403.02818, 2024

    arXiv 2024

  35. [35]

    Dgpinet- kd: Deep guided and progressive integration network with knowledge distillation for rgb-d indoor scene analysis,

    W. Zhou, B. Jian, M. Fang, X. Dong, Y . Liu, and Q. Jiang, “Dgpinet- kd: Deep guided and progressive integration network with knowledge distillation for rgb-d indoor scene analysis,”IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  36. [36]

    A comprehensive survey on hardware-aware neural architecture search,

    H. Benmeziane, K. E. Maghraoui, H. Ouarnoughi, S. Niar, M. Wistuba, and N. Wang, “A comprehensive survey on hardware-aware neural architecture search,”arXiv preprint arXiv:2101.09336, 2021

    Pith/arXiv arXiv 2021

  37. [37]

    Automl: A systematic review on automated machine learning with neural architecture search,

    I. Salehin, M. S. Islam, P. Saha, S. Noman, A. Tuni, M. M. Hasan, and M. A. Baten, “Automl: A systematic review on automated machine learning with neural architecture search,”Journal of Information and Intelligence, vol. 2, no. 1, pp. 52–81, 2024

    2024

  38. [38]

    Mnasnet: Platform-aware neural architecture search for mobile,

    M. Tan, B. Chen, R. Pang, V . Vasudevan, M. Sandler, A. Howard, and Q. V . Le, “Mnasnet: Platform-aware neural architecture search for mobile,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 2820–2828

    2019

  39. [39]

    Fast hardware- aware neural architecture search,

    L. L. Zhang, Y . Yang, Y . Jiang, W. Zhu, and Y . Liu, “Fast hardware- aware neural architecture search,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020, pp. 692–693

    2020

  40. [40]

    The deep learning compiler: A comprehensive survey,

    M. Li, Y . Liu, X. Liu, Q. Sun, X. You, H. Yang, Z. Luan, L. Gan, G. Yang, and D. Qian, “The deep learning compiler: A comprehensive survey,”IEEE Transactions on Parallel and Distributed Systems, vol. 32, no. 3, pp. 708–727, 2020

    2020

  41. [41]

    Hw-nas-bench: Hardware-aware neural architecture search benchmark,

    C. Li, Z. Yu, Y . Fu, Y . Zhang, Y . Zhao, H. You, Q. Yu, Y . Wang, and Y . Lin, “Hw-nas-bench: Hardware-aware neural architecture search benchmark,”arXiv preprint arXiv:2103.10584, 2021

    Pith/arXiv arXiv 2021

  42. [42]

    On- device training under 256kb memory,

    J. Lin, L. Zhu, W.-M. Chen, W.-C. Wang, C. Gan, and S. Han, “On- device training under 256kb memory,” inAnnual Conference on Neural Information Processing Systems (NeurIPS), 2022

    2022

  43. [43]

    Micronets: Neural network architectures for deploying tinyml applications on commodity microcontrollers,

    C. Banbury, C. Zhou, I. Fedorov, R. Matas, U. Thakker, D. Gope, V . Janapa Reddi, M. Mattina, and P. Whatmough, “Micronets: Neural network architectures for deploying tinyml applications on commodity microcontrollers,”Proceedings of Machine Learning and Systems, vol. 3, pp. 517–532, 2021

    2021

  44. [44]

    A machine learning-oriented survey on tiny machine learning,

    L. Capogrosso, F. Cunico, D. S. Cheng, F. Fummiet al., “A machine learning-oriented survey on tiny machine learning,”arXiv preprint arXiv:2309.11932, 2023

    Pith/arXiv arXiv 2023

  45. [45]

    Single path one-shot neural architecture search with uniform sampling,

    Z. Guo, X. Zhang, H. Mu, W. Heng, Z. Liu, Y . Wei, and J. Sun, “Single path one-shot neural architecture search with uniform sampling,” in European conference on computer vision. Springer, 2020, pp. 544– 560

    2020

  46. [46]

    Enhancing neural architecture search with multiple hardware constraints for deep learning model deployment on tiny iot devices,

    A. Burrello, M. Risso, B. A. Motetti, E. Macii, L. Benini, and D. J. Pagliari, “Enhancing neural architecture search with multiple hardware constraints for deep learning model deployment on tiny iot devices,” IEEE Transactions on Emerging Topics in Computing, 2023

    2023

  47. [47]

    Efficient evaluation methods for neural architecture search: A survey,

    X. Song, X. Xie, Z. Lv, G. G. Yen, W. Ding, J. Lv, and Y . Sun, “Efficient evaluation methods for neural architecture search: A survey,” IEEE Transactions on Artificial Intelligence, 2024

    2024

  48. [48]

    Zero-shot neural architecture search: Challenges, solutions, and opportunities,

    G. Li, D. Hoang, K. Bhardwaj, M. Lin, Z. Wang, and R. Mar- culescu, “Zero-shot neural architecture search: Challenges, solutions, and opportunities,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  49. [49]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,”Advances in Neural Information Processing Systems, vol. 34, pp. 12 077–12 090, 2021

    2021

  50. [50]

    A survey of multi- fingered robotic manipulation: Biological results, structural evolvements and learning methods,

    Y . Li, P. Wang, R. Li, M. Tao, Z. Liu, and H. Qiao, “A survey of multi- fingered robotic manipulation: Biological results, structural evolvements and learning methods,”Frontiers in Neurorobotics, p. 53, 2022

    2022