Pith. sign in

REVIEW 3 major objections 7 minor 1 cited by

A 1.3-million-parameter CNN can deliver high-quality promptable object segmentation inside a vision sensor in real time.

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

2026-07-14 22:31 UTC pith:MFZOVF5Y

load-bearing objection Solid IMX500 systems paper with real in-sensor ROI masks; Table I’s “beats SAM-H” numbers are protocol-mismatched and should not be taken at face value. the 3 major comments →

arxiv 2603.11917 v4 pith:MFZOVF5Y submitted 2026-03-12 cs.CV

PicoSAM3: Real-Time In-Sensor Region-of-Interest Segmentation

Pietro Bonazzi , Nicola Farronato , Stefan Zihlmann , Haotong Qin , Michele Magno This is my paper
classification cs.CV
keywords Edge AIIn-sensor computingPromptable segmentationKnowledge distillationQuantizationReal-time segmentationSegment Anything ModelSmart sensors
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.

This paper shows that promptable visual segmentation need not stay on large GPUs or even on a phone CPU: it can run fully inside a stacked smart camera sensor. PicoSAM3 is a dense convolutional network of about 1.3 million parameters that treats a hardware region-of-interest crop as an implicit box prompt, adds light channel attention and a dilated bottleneck, and is trained by distilling soft masks from the large SAM2 and SAM3 teachers. On standard COCO and LVIS ROI-crop tests it reaches 65.45 percent and 64.01 percent mean intersection-over-union, while the INT8 version fits in 1.31 MB and finishes inference in 11.82 ms on the Sony IMX500. Ablations attribute large accuracy gains to the distillation step and to flexible ROI prompting rather than older centered-point prompts. The practical claim is that privacy-preserving, low-latency segmentation for smart glasses and IoT cameras is already feasible without leaving the sensor package.

Core claim

High-quality, spatially flexible promptable segmentation is feasible directly at the sensor. PicoSAM3 (1.37 M parameters) combines a U-Net-style dense CNN, dilated bottleneck, Efficient Channel Attention, ROI-crop implicit prompting, and distillation from SAM2/SAM3 to obtain 65.45 percent mIoU on COCO and 64.01 percent on LVIS; the INT8 model retains essentially the same accuracy at 1.31 MB and 11.82 ms end-to-end latency on the IMX500 while obeying its memory and operator limits.

What carries the argument

Implicit ROI prompt encoding via centered object cropping (with padding) that is resized to 96 by 96, paired with offline knowledge distillation of temperature-scaled soft masks from SAM3 into a quantization-friendly CNN student. The crop itself supplies the spatial prior, so no extra prompt tensors are needed at inference and the pipeline maps directly onto the sensor’s hardware ROI mode.

Load-bearing premise

The claim that PicoSAM3 outperforms large foundation models rests on comparing numbers that may not have been measured under the same 96-by-96 ROI-crop protocol the tiny model actually uses.

What would settle it

Run SAM-H, FastSAM, TinySAM and EdgeSAM on exactly the same COCO and LVIS ROI crops resized to 96 by 96 that PicoSAM3 receives; if their mIoU then equals or exceeds PicoSAM3’s 65.45 percent / 64.01 percent, the superiority claim under in-sensor conditions fails.

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

If this is right

  • Promptable segmentation can execute fully in-sensor without cloud offload for latency- and privacy-sensitive devices.
  • Distillation from large SAM teachers can close most of the accuracy gap for sub-2 M-parameter CNNs operating on 96-by-96 ROI crops.
  • INT8 post-training quantization of this architecture incurs negligible mIoU loss while meeting extreme-edge memory budgets.
  • Flexible box/ROI prompting maps directly onto hardware ROI modes already present in intelligent vision sensors.
  • Hardware-aware dense CNNs avoid the unsupported operators and random memory access that block transformer SAM variants on the IMX500.

Where Pith is reading between the lines

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

  • The same ROI-crop plus distillation pattern could transfer to other stacked or always-on sensors that expose only RGB frames and a hardware crop API.
  • If large foundation models systematically collapse at 96-by-96 resolution, future edge work may need native low-resolution teachers rather than only compressing high-resolution ones.
  • Wearable AR pipelines could feed a tiny detector’s boxes straight into an in-sensor PicoSAM3-style mask head without ever leaving the camera package.
  • The adaptive weighting of teacher versus ground-truth loss by teacher confidence is a reusable recipe for distilling any large vision model onto extreme-edge students.

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 / 7 minor

Summary. PicoSAM3 is a 1.3–1.4 M-parameter fully convolutional, ROI-prompted segmentation student designed for the Sony IMX500’s <8 MB SRAM and restricted operator set. Prompts are encoded implicitly by centered padded square crops (no extra prompt channels); the architecture extends PicoSAM2 with a dilated bottleneck, ECA, and a refinement head; training uses adaptive distillation from SAM2/SAM3 soft masks plus ground-truth and area-preservation losses. On COCO/LVIS ROI-crop evaluation the FP32 model reports 65.45%/64.01% mIoU; INT8 PTQ yields 1.31 MB and 11.82 ms end-to-end latency on the IMX500 with <0.2% mIoU drop. Ablations (Table II) attribute large gains to ROI cropping and SAM3 distillation over supervised training.

Significance. If the in-sensor results hold under clarified evaluation, the paper is a concrete systems contribution for IEEE Sensors Journal: it shows that spatially flexible, box-style promptable segmentation can run fully on-sensor at ~84 FPS under real memory/operator constraints, with public code and measured silicon latency. Strengths include clean ablations separating ROI, teacher, architecture, and quantization; near-lossless INT8; and a hardware-aligned implicit-prompt design that maps to the IMX500 ROI API. The work is valuable even if headline comparisons to full-resolution foundation models are revised, because the deployment and PicoSAM2-to-PicoSAM3 progression are independently useful.

major comments (3)
  1. Table I / Abstract / §V.A: The central comparative claim that PicoSAM3 “surpasses SAM-H” by +11.85 mIoU on COCO (65.45% vs 53.6%) and similarly beats FastSAM/TinySAM/EdgeSAM/LiteSAM mixes incompatible protocols. PicoSAM3 is scored only on centered 96×96 ROI crops with implicit box prompts; the listed foundation/edge numbers are published full-image, high-resolution figures. §V.B and Fig. 3 themselves show large SAMs suffer severe feature collapse under the 96×96 crop regime (~25% mAP for SAM2.1 Large). Without a same-protocol re-evaluation column (or a clear split of “published full-res” vs “96×96 ROI-crop” scores), the “outperforming … at similar or lower complexity” narrative does not support superiority over prior art—only the narrower (still important) claim of feasible in-sensor deployment and gains over PicoSAM2 under matched conditions.
  2. §IV.A–B and evaluation protocol: “Promptable” is used throughout, but the system supports only single-object, box-like conditioning via hardware ROI crop (no point prompts, no multi-mask decoding, no iterative refinement at inference beyond re-cropping). Table I labels PicoSAM3 “Box” while several baselines are full SAM-style promptable models. Please define the supported prompt interface explicitly, report whether multi-object or point-prompt settings are out of scope, and ensure baseline comparisons use the same prompt type and crop protocol where claims of outperformance are made.
  3. §IV.D / Table I: Training is reported as one epoch on COCO with batch 64. For a distillation student claiming large gains over supervised training and SOTA among edge models, one epoch is unusually short and risks under-training or teacher-cache memorization. Provide training curves, multi-epoch ablations, or validation that longer training does not change the ranking vs PicoSAM2 and vs same-protocol baselines; otherwise the absolute mIoU numbers and distillation gains (+14.5% in the abstract) need stronger evidence of stability.
minor comments (7)
  1. §V: “PyTorch 3.7” is not a released version; correct to the actual version used.
  2. References: Milletari et al. V-Net appears twice ([49] and [50]); deduplicate.
  3. Table I: Q-PicoSAM3 size is listed as 1.30 MB in the table and 1.31 MB in the text; unify.
  4. Fig. 3 caption and §V.B: Clarify that “feature collapse” of large SAMs is under the authors’ 96×96 crop protocol so readers do not misread the Pareto plot as a general ranking of foundation models.
  5. Abstract vs body: Abstract says “1.3M parameters” and “distillation from SAM2 and SAM3”; body uses 1.37M and primarily reports SAM3 for the final model—align wording.
  6. Eq. (6)–(9): Temperature-scaled sigmoid is written σ_τ(x)=σ(τ·x); confirm whether τ multiplies logits before or after scaling as intended for “sharper” distributions (τ=5 typically softens if used as divisor).
  7. §III / Fig. 1: Minor grammar (“runs fully in-sensors”); polish for camera-ready.

Circularity Check

0 steps flagged

No derivation circularity: empirical distillation + held-out/hardware measurement; self-cites to PicoSAM2 are incremental prior work, not load-bearing identities.

full rationale

PicoSAM3 is a systems/engineering paper whose central claims (mIoU on COCO/LVIS ROI crops, INT8 size, 11.82 ms IMX500 latency) are obtained by training a CNN student against precomputed external SAM2/SAM3 teacher logits plus COCO ground-truth masks, then measuring on held-out splits and real silicon. No equation, definition, or fitted parameter forces the reported numbers by construction (Eqs. 1–9 define cropping, losses, and adaptive weighting; none equate output mIoU to an input). Self-citations to PicoSAM2 [5] appear in architecture description and Table I/II baselines as the immediate predecessor being improved; they supply neither a uniqueness theorem nor an ansatz that defines the new result. Protocol mismatches with published foundation-model numbers (acknowledged by the authors themselves in §V.B/Fig. 3) are a fairness/comparability issue, not circular derivation. The paper is therefore self-contained against external teachers, public benchmarks, and hardware; residual train/test dependence on COCO is ordinary empirical practice, not circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

Load-bearing content is mostly engineering choices and empirical hyperparameters, not new physical entities. The central feasibility claim rests on (a) the implicit-crop prompting axiom forced by RGB-only sensor I/O, (b) a handful of hand-set distillation/crop constants, and (c) the assumption that COCO-box ROI crops at 96×96 are a valid proxy for interactive in-sensor prompting. No new particles or forces; the invented entity is the model+prompting recipe itself.

free parameters (7)
  • ROI padding factor p = 0.1
    Hand-set to 0.1 on each side of the GT box before square crop; changes context and effective scale of every training/inference sample.
  • Network input resolution S = 96
    Fixed crop resize to 96×96; dominates accuracy–latency trade-off and is required for IMX500 memory fit.
  • Distillation temperature τ = 5
    Temperature-scaled sigmoid in L_teacher / L_gt; set to 5 to sharpen soft masks.
  • Area-preservation threshold ρ = 0.4
    L_area activates when predicted area < ρ × GT area; ρ=0.4 chosen to prevent collapse.
  • L_area loss weight = 0.4
    Fixed coefficient 0.4 in L_total.
  • AdamW learning rate and schedule = 3e-4, 1 epoch, bs=64
    3e-4 with 1000-step linear warmup; one-epoch COCO training only.
  • Architecture channel widths / dilation = 1.37M params design
    Encoder 48→96→160→256, bottleneck 320, dilation=2, ECA, refinement head—hand-designed for IMX500 operator set and size budget.
axioms (5)
  • ad hoc to paper RGB-only IMX500 inputs forbid explicit prompt tensors; a centered padded square crop is a sufficient implicit encoding of box/point prompts for promptable segmentation.
    Stated in §IV.A as the enabling design for sensor deployment; if false, the model is only a crop segmenter, not promptable SAM-style.
  • domain assumption Soft masks from SAM2/SAM3 on the same boxes are valid supervision that transfers to a tiny CNN student under the adaptive α confidence weighting.
    Core of §IV.C distillation pipeline; ablations support gains but do not prove teacher correctness on every crop.
  • domain assumption COCO/LVIS instance boxes with 10% pad, square max-side crop, and resize to 96×96 are a fair proxy for interactive ROI prompting on-device.
    Evaluation and training both use this pipeline (§IV, §V); real user ROIs may differ in tightness and clutter.
  • domain assumption Depthwise-separable CNN + ECA + dilated bottleneck can approximate the needed spatial selectivity without self-attention under INT8 DSP constraints.
    Architectural thesis of §IV.B and related-work contrast with transformer students.
  • standard math Standard segmentation metrics (mIoU, mAP) and PTQ calibration on 10 COCO val batches suffice to claim negligible quantization loss and real-time feasibility.
    Conventional CV evaluation and Sony MCT PTQ practice as used in §IV.D–§V.
invented entities (2)
  • PicoSAM3 (ROI-implicit CNN student + SAM3 distillation recipe) independent evidence
    purpose: Provide a quantization-friendly, IMX500-deployable promptable segmenter in the ~1.3M parameter regime.
    Named model and training/deployment stack introduced by the paper; independent evidence is the public code and reported silicon latency, not an external physical discovery.
  • Implicit prompt encoding via centered ROI cropping no independent evidence
    purpose: Replace explicit prompt channels unsupported by the sensor with a spatial prior aligned to hardware ROI APIs.
    Design pattern specialized for this hardware constraint; falsifiable by showing failure under off-center or multi-object ROIs, which the paper only partially explores.

pith-pipeline@v1.1.0-grok45 · 18933 in / 4102 out tokens · 39337 ms · 2026-07-14T22:31:37.101308+00:00 · methodology

0 comments
read the original abstract

Real-time, on-device segmentation is critical for latency-sensitive and privacy-aware applications such as smart glasses and Internet-of-Things devices. We introduce PicoSAM3, a lightweight promptable visual segmentation model optimized for edge and in-sensor execution, including deployment on the Sony IMX500 vision sensor. PicoSAM3 has 1.3M parameters and combines a dense CNN architecture with region of interest prompt encoding, Efficient Channel Attention, and knowledge distillation from SAM2 and SAM3. On COCO and LVIS, PicoSAM3 achieves 65.45% and 64.01% mIoU, respectively, outperforming existing SAM-based and edge-oriented baselines at similar or lower complexity. The INT8 quantized model preserves accuracy with negligible degradation while enabling real-time in-sensor inference at 11.82ms latency on the IMX500, fully complying with its memory and operator constraints. Ablation studies show that distillation from large SAM models yields up to +14.5% mIoU improvement over supervised training and demonstrate that high-quality, spatially flexible promptable segmentation is feasible directly at the sensor level.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Exploiting In-Sensor Computing for Energy-Efficient Earth Observation

    cs.CV 2026-05 unverdicted novelty 4.0

    TinyML models on IMX500 deliver 96.68% accuracy on EuroSAT at 17.4 FPS and 14.19 mJ per inference within 8 MB memory for in-sensor EO.

Reference graph

Works this paper leans on

57 extracted references · 15 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll´ar, and R. Girshick, “Segment anything,”IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    2023

  2. [2]

    Sam2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y . Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C. Wu, R. Girshick, P. Doll ´ar, and C. Feichtenhofer, “Sam2: Segment anything in images and videos,”arXiv, 2408.00714, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Sam 3: Segment anything with concepts,

    N. Carion, L. Gustafson, Y .-T. Hu, S. Debnath, R. Hu, D. Suris, C. Ryali, K. V . Alwala, H. Khedr, A. Huang, J. Lei, T. Ma, B. Guo, A. Kalla, M. Marks, J. Greer, M. Wang, P. Sun, R. R¨adle, T. Afouras, E. Mavroudi, K. Xu, T.-H. Wu, Y . Zhou, L. Momeni, R. Hazra, S. Ding, S. Vaze, F. Porcher, F. Li, S. Li, A. Kamath, H. K. Cheng, P. Doll ´ar, N. Ravi, K. ...

    Pith/arXiv arXiv 2025

  4. [4]

    Efficient remote sensing image target detection network with shape-location awareness enhance- ments,

    F. Fan, M. Zhang, D. Yu, J. Li, and G. Liu, “Efficient remote sensing image target detection network with shape-location awareness enhance- ments,”IEEE Sensors Journal, 2024

    2024

  5. [5]

    Pi- cosam2: Low-latency segmentation in-sensor for edge vision applica- tions,

    P. Bonazzi, N. Farronato, S. Zihlmann, H. Qin, and M. Magno, “Pi- cosam2: Low-latency segmentation in-sensor for edge vision applica- tions,”IEEE Sensors Conference, 2025

    2025

  6. [6]

    A novel embedded deep learning wearable sensor for fall detection,

    S. Campanella, A. Alnasef, L. Falaschetti, A. Belli, P. Pierleoni, and L. Palma, “A novel embedded deep learning wearable sensor for fall detection,”IEEE Sensors Journal, 2024

    2024

  7. [7]

    Low-power detection and classification for in-sensor predictive main- tenance based on vibration monitoring,

    P. Vitolo, A. De Vita, L. D. Benedetto, D. Pau, and G. D. Licciardo, “Low-power detection and classification for in-sensor predictive main- tenance based on vibration monitoring,”IEEE Sensors Journal, 2022

    2022

  8. [8]

    Ultra-efficient on-device object detection on ai-integrated smart glasses,

    J. Moosmann, P. Bonazzi, Y . Li, S. Bian, P. Mayer, L. Benini, and M. Magno, “Ultra-efficient on-device object detection on ai-integrated smart glasses,”IEEE/CVF European Conference on Computer Vision (ECCV), 2023

    2023

  9. [9]

    Fann-on-mcu: An open-source toolkit for energy-efficient neural network inference at the edge of the internet of things,

    X. Wang, M. Magno, L. Cavigelli, and L. Benini, “Fann-on-mcu: An open-source toolkit for energy-efficient neural network inference at the edge of the internet of things,”IEEE Internet of Things Journal, 2020

    2020

  10. [10]

    Survey and comparison of milliwatts micro controllers for tiny machine learning at the edge,

    M. Giordano, L. Piccinelli, and M. Magno, “Survey and comparison of milliwatts micro controllers for tiny machine learning at the edge,”IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS), 2022

    2022

  11. [11]

    Low latency visual inertial odom- etry with on-sensor accelerated optical flow for resource-constrained uavs,

    J. K ¨uhne, M. Magno, and L. Benini, “Low latency visual inertial odom- etry with on-sensor accelerated optical flow for resource-constrained uavs,”IEEE Sensors Journal, 2025

    2025

  12. [12]

    A real-time intelligent system based on machine-learning methods for improving communication in sign language,

    V . Leiva, M. Z. U. Rahman, M. A. Akbar, C. Castro, M. Huerta, and M. T. Riaz, “A real-time intelligent system based on machine-learning methods for improving communication in sign language,”IEEE Access, 2025

    2025

  13. [13]

    Design and implementation of a resnet-lstm-ghost architecture for gas concentration estimation of electronic noses,

    G. Wei, X. Liu, A. He, W. Zhang, S. Jiao, and B. Wang, “Design and implementation of a resnet-lstm-ghost architecture for gas concentration estimation of electronic noses,”IEEE Sensors Journal, 2024

    2024

  14. [14]

    A mixture-gas edge-computing multisensor device with generative learning framework,

    J. Cho, Y . J. Pyeon, Y . M. Kwon, Y . Kim, J. Yeom, M. W. Kim, C. S. Park, I.-H. Kim, Y . Lee, and J. J. Kim, “A mixture-gas edge-computing multisensor device with generative learning framework,”IEEE Sensors Journal, 2024

    2024

  15. [15]

    Aviear: An iot-based low-power solution for acoustic monitoring of avian species,

    R. Verma and S. Kumar, “Aviear: An iot-based low-power solution for acoustic monitoring of avian species,”IEEE Sensors Journal, 2024

    2024

  16. [16]

    Sim- ulation, design, and application of intelligent-edge-based soft magnetic tactile sensor with super-resolution,

    Y . Zhou, Y . Luo, Z. Yan, Y . Jin, S. Jiang, Z. Wang, and B. He, “Sim- ulation, design, and application of intelligent-edge-based soft magnetic tactile sensor with super-resolution,”IEEE Sensors Journal, 2024

    2024

  17. [17]

    Design and implemen- tation of an arm-based ai module for ectopic beat classification using custom and structural pruned lightweight cnn,

    Y .-L. Xie, X.-R. Lin, C.-Y . Lee, and C.-W. Lin, “Design and implemen- tation of an arm-based ai module for ectopic beat classification using custom and structural pruned lightweight cnn,”IEEE Sensors Journal, 2024

    2024

  18. [18]

    Preliminary analysis of the exploitation of qvar sensor for gesture recognition,

    A. Messina, A. Lazzaro, R. Carotenuto, and M. Merenda, “Preliminary analysis of the exploitation of qvar sensor for gesture recognition,”Inter- national Conference on Smart and Sustainable Technologies (SpliTech), 2025

    2025

  19. [19]

    Sony imx500,

    “Sony imx500,” 2023. [Online]. Available: https://developer.sony.com/ imx500/

    2023

  20. [20]

    A 1/2.3inch 12.3mpixel with on-chip 4.97tops/w cnn processor back- illuminated stacked cmos image sensor,

    R. Eki, S. Yamada, H. Ozawa, H. Kai, K. Okuike, H. Gowtham, H. Nakanishi, E. Almog, Y . Livne, G. Yuval, E. Zyss, and T. Izawa, “A 1/2.3inch 12.3mpixel with on-chip 4.97tops/w cnn processor back- illuminated stacked cmos image sensor,”IEEE International Solid- State Circuits Conference (ISSCC), 2021

    2021

  21. [21]

    Tinytracker: Ultra- fast and ultra-low-power edge vision for in-sensor gaze estimation,

    P. Bonazzi, T. R ¨uegg, S. Bian, Y . Li, and M. Magno, “Tinytracker: Ultra- fast and ultra-low-power edge vision for in-sensor gaze estimation,” IEEE Sensors Conference, 2023

    2023

  22. [22]

    Near-sensor and in-sensor computing,

    F. Zhou and Y . Chai, “Near-sensor and in-sensor computing,”Nature Electronics, 2020

    2020

  23. [23]

    Tinysam: Pushing the envelope for efficient segment anything model,

    H. Shu, W. Li, Y . Tang, Y . Zhang, Y . Chen, H. Li, Y . Wang, and X. Chen, “Tinysam: Pushing the envelope for efficient segment anything model,”arXiv preprint arXiv:2312.13789, 2023, also available via AAAI implementation

    Pith/arXiv arXiv 2023

  24. [24]

    Edgesam: Prompt-in-the-loop distillation for on-device deployment of sam,

    C. Zhou, X. Li, C. C. Loy, and B. Dai, “Edgesam: Prompt-in-the-loop distillation for on-device deployment of sam,”arXiv, 2312.06660, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Mo- bilesamv2: Faster segment anything to everything,

    C. Zhang, D. Han, S. Zheng, J. Choi, T.-H. Kim, and C. S. Hong, “Mo- bilesamv2: Faster segment anything to everything,”arXiv, 2312.09579, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Lite-sam is actually what you need for segment everything,

    J. Fu, Y . Yu, N. Li, Y . Zhang, Q. Chen, J. Xiong, J. Yin, and Z. Xiang, “Lite-sam is actually what you need for segment everything,”European Conference on Computer Vision (ECCV), 2024

    2024

  27. [27]

    Efficient deep learning: A survey on making deep learning models smaller, faster, and better,

    G. Menghani, “Efficient deep learning: A survey on making deep learning models smaller, faster, and better,”ACM Computing Surveys, 2023

    2023

  28. [28]

    Mobilenetv4: Universal models for the mobile ecosystem,

    D. Qin, C. Leichner, M. Delakis, M. Fornoni, S. Luo, F. Yang, W. Wang, C. Banbury, C. Ye, B. Akinet al., “Mobilenetv4: Universal models for the mobile ecosystem,”European Conference on Computer Vision (ECCV), 2024

    2024

  29. [29]

    Eca-net: Efficient channel attention for deep convolutional neural networks,

    Q. Wang, B. Wu, P. Zhu, P. Li, W. Zuo, and Q. Hu, “Eca-net: Efficient channel attention for deep convolutional neural networks,”IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  30. [30]

    A survey on segment anything model (sam): Vision foundation model meets prompt engineering,

    C. Zhang, J. Cho, F. D. Puspitasari, S. Zheng, C. Li, Y . Qiao, T. Kang, X. Shan, C. Zhang, C. Qinet al., “A survey on segment anything model (sam): Vision foundation model meets prompt engineering,” arXiv, 2306.06211, 2023. BONAZZIet al.: : ARTICLE UNDER REVIEW A T IEEE SENSORS JOURNAL 9

    Pith/arXiv arXiv 2023

  31. [31]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv, 2010.11929, 2020

    Pith/arXiv arXiv 2010

  32. [32]

    Fast segment anything,

    X. Zhao, W. Ding, Y . An, Y . Du, T. Yu, M. Li, M. Tang, and J. Wang, “Fast segment anything,” 2023

    2023

  33. [33]

    Faster segment anything: Towards lightweight sam for mobile applications,

    C. Zhang, D. Han, Y . Qiao, J. U. Kim, S.-H. Bae, S. Lee, and C. S. Hong, “Faster segment anything: Towards lightweight sam for mobile applications,”arXiv, 2306.14289, 2023

    Pith/arXiv arXiv 2023

  34. [34]

    Expediting large-scale vision transformer for dense prediction without fine-tuning,

    W. Liang, Y . Yuan, H. Ding, X. Luo, W. Lin, D. Jia, Z. Zhang, C. Zhang, and H. Hu, “Expediting large-scale vision transformer for dense prediction without fine-tuning,”Advances in Neural Information Processing Systems, 2022

    2022

  35. [35]

    0.1% data makes segment anything slim,

    Z. Chen, G. Fang, X. Ma, and X. Wang, “0.1% data makes segment anything slim,”arXiv, 2312.05284, 2023

    Pith/arXiv arXiv 2023

  36. [36]

    Ptq4sam: Post-training quantization for segment anything,

    C. Lv, H. Chen, J. Guo, Y . Ding, and X. Liu, “Ptq4sam: Post-training quantization for segment anything,”IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  37. [37]

    Pq-sam: Post-training quantization for segment anything model,

    X. Liu, X. Ding, L. Yu, Y . Xi, W. Li, Z. Tu, J. Hu, H. Chen, B. Yin, and Z. Xiong, “Pq-sam: Post-training quantization for segment anything model,”European Conference on Computer Vision (ECCV), 2024

    2024

  38. [38]

    Post training 4-bit quantization of convolutional networks for rapid-deployment,

    R. Banner, I. Hubara, E. Hoffer, and D. Soudry, “Post training 4-bit quantization of convolutional networks for rapid-deployment,”Advances in Neural Information Processing Systems (NeurIPS), 2019

    2019

  39. [39]

    Up or down? adaptive rounding for post-training quantization,

    M. Nagel, M. van Baalen, T. Blankevoort, and M. Welling, “Up or down? adaptive rounding for post-training quantization,”European Conference on Computer Vision (ECCV), 2020

    2020

  40. [40]

    Hiera: A hierarchical vision transformer without the bells-and-whistles,

    C. Ryali, Y .-T. Hu, D. Bolya, C. Wei, H. Fan, P.-Y . Huang, V . Aggarwal, A. Chowdhury, O. Poursaeed, J. Hoffmanet al., “Hiera: A hierarchical vision transformer without the bells-and-whistles,”International Con- ference on Machine Learning (ICML), 2023

    2023

  41. [41]

    Masked autoencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,”IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  42. [42]

    Autoppn: Learning to design promptable pyramid networks for efficient segmentation,

    Y . Xu, F. Yu, and B. Zhou, “Autoppn: Learning to design promptable pyramid networks for efficient segmentation,”IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    2023

  43. [43]

    Litevit: Towards efficient vision transformers with enhanced token mixing,

    H. Wang, K. Wang, Y . Xu, C. Xu, and C. Shen, “Litevit: Towards efficient vision transformers with enhanced token mixing,”arXiv, 2303.13429, 2023

    Pith/arXiv arXiv 2023

  44. [44]

    Edge AI-enabled chicken health detection based on enhanced FCOS-Lite and knowledge distillation,

    Q. Tong, J. Wang, W. Yang, S. Wu, W. Zhang, C. Sun, and K. Xu, “Edge AI-enabled chicken health detection based on enhanced FCOS-Lite and knowledge distillation,”Computers and Electronics in Agriculture, vol. 226, p. 109432, 2024

    2024

  45. [45]

    Pedestrian Warning: Intelligent Vision Sensor vs. Edge AI with LTE C-V2X in a Smart City,

    T. Cui, Z. Zhang, C. Sun, S. Wang, H. Li, and W. Zhang, “Pedestrian Warning: Intelligent Vision Sensor vs. Edge AI with LTE C-V2X in a Smart City,” inIEEE 99th Vehicular Technology Conference (VTC2024- Spring), 2024

    2024

  46. [46]

    U-Net: Convolutional net- works for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional net- works for biomedical image segmentation,” 2015

    2015

  47. [47]

    Xception: Deep learning with depthwise separable convolu- tions,

    F. Chollet, “Xception: Deep learning with depthwise separable convolu- tions,”IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  48. [48]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,”arXiv, 1503.02531, 2015

    Pith/arXiv arXiv 2015

  49. [49]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation,

    F. Milletari, N. Navab, and S.-A. Ahmadi, “V-net: Fully convolutional neural networks for volumetric medical image segmentation,”Interna- tional Conference on 3D Vision (3DV), 2016

    2016

  50. [50]

    V-net: Fully convolutional neu- ral networks for volumetric medical image segmentation,

    F. Milletari, N. Navab, and S. Ahmadi, “V-net: Fully convolutional neu- ral networks for volumetric medical image segmentation,”International Conference on 3D Vision (3DV), 2016

    2016

  51. [51]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,

    V . Sanh, L. Debut, J. Chaumond, and T. Wolf, “Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,”arXiv, 1910.01108, 2019

    Pith/arXiv arXiv 1910

  52. [52]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv, 2019

    2019

  53. [53]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshick, J. Hays, P. Perona, D. Ramanan, C. L. Zitnick, and P. Doll ´ar, “Microsoft coco: Common objects in context,”arXiv, 1405.0312, 2014

    Pith/arXiv arXiv 2014

  54. [54]

    Hptq: Hardware-friendly post training quantization,

    H. V . Habi, R. Peretz, E. Cohen, L. Dikstein, O. Dror, I. Diamant, R. H. Jennings, and A. Netzer, “Hptq: Hardware-friendly post training quantization,”arXiv, 2109.09113, 2021

    Pith/arXiv arXiv 2021

  55. [55]

    Eptq: Enhanced post- training quantization via hessian-guided network-wise optimization,

    O. Gordon, E. Cohen, H. V . Habi, and A. Netzer, “Eptq: Enhanced post- training quantization via hessian-guided network-wise optimization,” in European Conference on Computer Vision (ECCV) Workshops, 2024

    2024

  56. [56]

    Data generation for hardware-friendly post-training quantization,

    L. Dikstein, A. Lapid, A. Netzer, and H. V . Habi, “Data generation for hardware-friendly post-training quantization,” inIEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025

    2025

  57. [57]

    A survey of quantization methods for efficient neural network inference,

    A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, and K. Keutzer, “A survey of quantization methods for efficient neural network inference,” 2021. [Online]. Available: https://arxiv.org/abs/2103.13630

    Pith/arXiv arXiv 2021