Pith. sign in

REVIEW 4 major objections 4 minor 49 references

TinyGLASS: Real-Time Self-Supervised In-Sensor Anomaly Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read TinyGLASS demonstrates real-time self-supervised anomaly detection on a commercial in-sensor processor.

desk verdict Real in-sensor deployment numbers worth taking seriously, but the headline AUROC may be a selection artifact; fix the validation protocol and this becomes a useful systems paper. read the letter →

arxiv 2603.16451 v3 pith:WXFDP2CZ submitted 2026-03-17 cs.CV

Pietro Bonazzi , Rafael Sutter , Luigi Capogrosso , Mischa Buob , Michele Magno This is my paper
classification cs.CV
keywords TinyGLASSin-sensoranomalydetectionself-supervisedlearningindustrialdefectINT8quantizationSonyIMX500MVTec-ADedgeAI
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

TinyGLASS aims to show that self-supervised visual anomaly detection—learning to flag defects from defect-free images alone—can run in real time directly on a commercial image sensor. The authors compress the GLASS pipeline 8.6-fold by replacing its WideResNet-50 backbone with a truncated ResNet-18 and re-engineering the forward pass for static tracing and INT8 quantization. They report 94.2% image-level AUROC on MVTec-AD, 20 FPS throughput and 4.0 mJ per inference on the Sony IMX500, staying within the 8 MB on-sensor memory limit. If correct, this is the first real-time in-sensor visual anomaly detection system, a step toward low-cost, privacy-preserving quality control in factories.

What carries the argument

The load-bearing object is the TinyGLASS network: a truncated ResNet-18 (only layers up to layer 3) that emits 384-dimensional concatenated embeddings, a PatchMaker that constructs multi-scale local and global views, and a discriminator that outputs a dense anomaly heatmap. The paper's enabling move is the deployment-oriented rewrite of GLASS: replacing tuple returns and dynamic reshapes with a single static 4D tensor pipeline, which allows the whole model to be traced and quantized to INT8. This transformation is what lets the model fit in 8 MB and execute on the IMX500's in-sensor CNN processor.

What would settle it

Retrain TinyGLASS on MVTec-AD with an explicit validation split for checkpoint selection, then evaluate on the official test set; if the resulting image-level AUROC is materially lower than 94.2%, the published number is an artifact of selection on test labels.

Watch

Extended reading notes

Core claim

The paper claims that GLASS's anomaly detection capability—synthesizing 'fake' anomalies via global and local feature perturbations and training a discriminator to separate normal from abnormal patches—can be preserved in a 2.9 M-parameter model that runs entirely on a commercial image sensor's neural network accelerator. By truncating a ResNet-18 to layer 3, concatenating multi-scale features into a single static 4D tensor, and reworking the discriminator input path to avoid dynamic shapes, the authors enable static graph tracing and INT8 quantization. The deployed TinyGLASS outputs patch-level anomaly heatmaps directly on the Sony IMX500 at 20 FPS, consuming 4.0 mJ per inference, while ach

Load-bearing premise

The reported 94.2% AUROC hinges on the unstated assumption that the best checkpoint was selected without peeking at the test set; the paper does not define a validation split for model selection.

Editorial extensions

If this is right

  • Self-supervised anomaly detection can be executed entirely on a camera sensor, meaning only compact anomaly heatmaps—not full images—leave the chip, slashing bandwidth and latency for inspection systems.
  • An 8.6x parameter reduction costs roughly 4.9 points of image-level AUROC on MVTec-AD, a trade-off that will often be acceptable for real-time edge inspection.
  • The quantized model's 20 FPS and 4.0 mJ per inference suggest continuous in-line quality control is feasible on low-power embedded devices.
  • Robustness experiments indicate that up to 30% contamination of the training data degrades I-AUROC only moderately, so the method tolerates imperfect data collection.
  • The new MMS dataset, captured with both a microscope and the IMX500, provides a benchmark for cross-device anomaly detection.

Reading between the lines

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

  • The same static-4D-tensor/INT8 recipe could be applied to other embedding-based anomaly detectors (e.g., SimpleNet-style discriminators), potentially bringing memory-bank and reconstruction methods onto in-sensor processors as well.
  • Because only the heatmap is output by the sensor, a host could fuse anomaly maps from multiple sensors or over time without transferring raw video, enabling multi-view inspection at very low bandwidth.
  • A natural next experiment is to measure sustained performance on a moving production line with the IMX500, since the paper's controlled lab conditions may not capture vibration, lighting drift, or thermal throttling.
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 / 4 minor

Summary. TinyGLASS adapts the GLASS anomaly-detection framework to a ResNet-18 backbone with static graph tracing and INT8 quantization, targeting the Sony IMX500 in-sensor processor. The paper reports an 8.6x parameter reduction, 94.2% image-level and 90.9% pixel-level AUROC on MVTec-AD, 88.9% I-AUROC on a newly introduced MMS industrial dataset, and a deployed system running at 20 FPS, 4.0 mJ per inference, and 470 GMAC/J on the IMX500. It also studies robustness to contaminated training data and introduces the MMS dataset with microscope and IMX500 images. The central claim is that this is the first real-time in-sensor visual anomaly detection demonstration on a commercial intelligent vision sensor.

Significance. If the accuracy and deployment numbers hold under a sound evaluation protocol, this is a useful systems contribution: it shows that a reasonably compact anomaly-detection model can run inside a commercial sensor at real-time throughput with low energy, and it provides a new industrial defect dataset for cross-device study. The energy and throughput figures are internally consistent (1.88 GMAC per inference, 20 FPS, 4.0 mJ, 470 GMAC/J) and the 8.6x parameter reduction is clearly quantified. However, the accuracy evaluation currently lacks a defined validation split, per-class results, error bars, and baselines on the new dataset, so the strength of the central accuracy claim cannot be assessed as reported.

major comments (4)
  1. [III-C and Section IV] The training protocol states that 'the best model is selected on the basis of the image-level AUROC performance' but no validation split is defined. Since TinyGLASS is trained only on normal images, image-level AUROC requires ground-truth labels of anomalous images, which exist only in the test splits of MVTec-AD and MMS. If selection is performed on these test labels during the 200 training epochs, the reported 94.2% and 88.9% I-AUROC values are selection artifacts (the maximum over checkpoints on the test distribution), not unbiased estimates. This directly affects the paper's central claim of competitive detection performance and the 4.9-pt drop versus GLASS. Please define a held-out validation protocol (e.g., a separate split or a fixed training schedule with no test-set access) and report the selected epoch and the resulting performance on the true test set, ideally with multiple se
  2. [Table II] The MVTec-AD comparison reports only mean image-level and pixel-level AUROC over the 15 categories, with no per-class breakdown and no standard deviation or number of runs. A 94.2% mean could hide large per-category failures, and the claim that the 8.6x compression gives 'competitive' performance cannot be evaluated without per-class results. Please include per-class AUROC tables and confidence intervals (or at least standard deviations) for both float-32 and int-8 models.
  3. [Section IV and Figure 3] The robustness claim of 'stable performance under moderate levels of training data contamination' is not supported by the evidence as presented. On MVTec-AD carpet, pixel-level AUROC drops from 99.2% at 0% contamination to 83.3% at 5% and 80.8% at 30% — a 16-point drop at only 5% contamination. The experiment is also limited to one MVTec category. Please define a stability criterion, report results across multiple categories or with error bars, and discuss whether the observed P-AUROC drop is acceptable for the target application.
  4. [III-D and Section IV] The MMS dataset is introduced with two camera setups (microscope and IMX500) and the paper claims 'cross-device evaluation', but the reported 88.9% I-AUROC is for microscope images only. No results on the IMX500-captured subset are given, no baseline method (e.g., GLASS, SimpleNet, or PatchCore) is evaluated on MMS, and the 'Test+Val' column in Table I is not defined with a concrete split procedure. Please provide results on the IMX500 subset, a baseline comparison on the same splits, and a clear description of how train/test splits are formed.
minor comments (4)
  1. [Figure 1 and Section V] Figure 1 states '5.41 MB of total memory' while Section V refers to the '8 MB memory constraint of the target platform'. Clarify whether 5.41 MB is the quantized model size or the runtime memory footprint, and report the int-8 parameter count and model size in Table II (currently the int-8 row has '-').
  2. [Equation (1)] The focal loss hyperparameters (alpha, gamma) are not defined. Please specify them for reproducibility.
  3. [Table I] The MMS table combines 'Test+Val'; specify whether any validation subset is used, how it is split, and whether the model-selection issue in III-C also applies to MMS.
  4. [Section IV] The runtime claim 'approximately 20 FPS' could be reported with the measured mean and variance over a fixed number of inference iterations, as is standard for hardware profiling.

Circularity Check

0 steps flagged · score 0.0 of 10

No meaningful circularity: TinyGLASS's claims rest on external benchmarks and deployment measurements; the AUROC-based checkpoint selection is a soundness concern, not a circular step.

full rationale

The paper is an engineering adaptation of an external method (GLASS, Chen et al., ECCV 2024) rather than a derivation that defines its outputs in terms of its inputs. The core claims—8.6x parameter compression, 94.2% I-AUROC on MVTec-AD, 20 FPS on IMX500—are empirical results evaluated against standard external benchmarks and a new dataset, not consequences of a self-citation chain or of fitting parameters to the predicted quantity. Self-citations appear (e.g., Bonazzi et al., Capogrosso et al.) but only as related-work context or platform references; none is load-bearing for the central result. The one flagged issue, 'The best model is selected on the basis of the image-level AUROC performance' (Section III-C) with no validation split defined, is a potential test-set selection bias / soundness concern: if AUROC is computed on test labels during checkpoint selection, the reported number may be optimistic. However, this does not make the derivation circular by construction—the AUROC is a selection criterion over training checkpoints, not a fitted parameter renamed as a prediction, and the deployment claims (FPS, memory, energy) are independent of this selection. Therefore no circular step is present, and the paper should be scored 0 for circularity while the validation-protocol ambiguity is noted as a correctness/reproducibility risk.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The reported results are empirical system measurements, not a derivation. The central accuracy claim depends on standard training hyperparameters, ImageNet-pretrained features, the GLASS loss transferring to a smaller backbone, and an unspecified patch-to-image aggregation. There is no code or dataset release, so the MMS results are not independently checkable.

free parameters (5)
  • Learning rates = 1e-4 (feature extractor), 2e-4 (discriminator)
    Chosen without ablation; they control the optimization of the whole pipeline and affect the final AUROC.
  • Training length and batch size = 200 epochs, batch size 8
    Maximum epochs with best-model selection by AUROC; no early-stopping or validation split is described.
  • Augmentation probability = 0.5
    Random rotations, translations, color jitter, and flips applied with probability 0.5; no sensitivity analysis.
  • Input preprocessing = resize to 256x256, ImageNet normalization
    Standard choice that affects feature statistics and the reported 1.88 GMAC count.
  • Model selection epoch = argmax image-level AUROC
    If the AUROC used for selection is computed on test labels, this is a form of fitting to the benchmark; no independent validation set is specified.
assumptions (4)
  • domain assumption ImageNet-pretrained ResNet-18 features transfer to industrial defect detection.
    Images are normalized with ImageNet statistics and the backbone is ResNet-18; pre-training is implied by the method lineage but not explicitly stated. No from-scratch baseline is given. Location: Section III-B/III-D.
  • domain assumption GLASS's GAS/LAS objective remains effective when the backbone is changed from WideResNet-50 to ResNet-18.
    The paper changes the backbone but keeps the Eq. (1) loss and GLASS synthesis; no ablation isolates the backbone effect. Location: Section III-B/C.
  • domain assumption Patch-level anomaly scores can be aggregated into a valid image-level score by an unspecified method.
    The text only says patch scores 'are aggregated to produce final image-level predictions'; the aggregation rule is load-bearing for all I-AUROC numbers. Location: Section III-C.
  • ad hoc to paper MVTec-AD and MMS test labels are not used during model selection.
    The paper selects the best model by AUROC but does not define a validation set; if this axiom is false, the reported AUROCs are optimistically selected. Location: Section III-C.
invented entities (1)
  • MMS Dataset
    purpose: Custom industrial micro-component defect dataset (normal, crack-hole, scratch, half) captured with a microscope and the IMX500 sensor for cross-device evaluation.
    No download URL or release statement is provided, so all MMS results cannot be independently reproduced without the authors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TinyGLASS: Real-Time Self-Supervised In-Sensor Anomaly Detection." pith.science (2026). https://pith.science/paper/WXFDP2CZ

@misc{pith2026260316451,
  author       = {Pith},
  title        = {Pith review of: TinyGLASS: Real-Time Self-Supervised In-Sensor Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WXFDP2CZ}},
  note         = {Machine review of arXiv:2603.16451}
}
read the original abstract

Anomaly detection plays a key role in industrial quality control, where defects must be identified despite the scarcity of labeled faulty samples. Recent self-supervised approaches, such as GLASS, learn normal visual patterns using only defect-free data and have shown strong performance on industrial benchmarks. However, their computational requirements limit their deployment on resource-constrained edge platforms, especially within in-sensor processing architectures. This work introduces TinyGLASS, a lightweight adaptation of the GLASS framework designed for real-time edge and in-sensor anomaly detection. The proposed architecture replaces the original WideResNet-50 backbone with a compact ResNet-18 and introduces deployment-based modifications that enable static graph tracing and INT8 quantization. We evaluate the proposed approach on the Sony IMX500 intelligent vision sensor, exploiting the in-sensor processor using the Sony Model Compression Toolkit. In addition to evaluating performance on the MVTec-AD benchmark, we investigate robustness to contaminated training data and introduce a custom industrial dataset, named MMS Dataset, for cross-device evaluation. Experimental results show that TinyGLASS achieves 8.6x parameter compression while maintaining competitive detection performance, reaching 94.2% image-level AUROC on MVTec-AD and operating at 20 FPS within the 8 MB memory constraints of the IMX500 platform. System profiling showcases low power consumption (4.0 mJ per inference), real-time end-to-end throughput (20 FPS), and high energy efficiency (470 GMAC/J). Furthermore, the model demonstrates stable performance under moderate levels of training data contamination.

Figures

Figures reproduced from arXiv: 2603.16451 by the authors.

Figure 1
Figure 1. Schematic of the proposed TinyGLASS architecture for end-to-end [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. MMS Dataset. Left: stereo microscope acquisition setup. Right: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 1 linked inside Pith

  1. [1]

    Self-supervised anomaly detection in computer vision and beyond: A survey and outlook,

    H. Hojjatiet al., “Self-supervised anomaly detection in computer vision and beyond: A survey and outlook,”Neural Networks, vol. 172, p. 106106, 2024

    2024

  2. [2]

    A survey of deep learning for industrial visual anomaly detection,

    Z. Liet al., “A survey of deep learning for industrial visual anomaly detection,”Artificial Intelligence Review, vol. 58, no. 9, 2025

    2025

  3. [3]

    Deep Industrial Image Anomaly Detection: A Survey,

    J. Liuet al., “Deep Industrial Image Anomaly Detection: A Survey,” Machine Intelligence Research, vol. 21, no. 1, pp. 104–135, 2024

    2024

  4. [4]

    An Adaptable and Unsupervised TinyML Anomaly Detection System for Extreme Industrial Environments,

    M. Antoniniet al., “An Adaptable and Unsupervised TinyML Anomaly Detection System for Extreme Industrial Environments,”Sensors, vol. 23, no. 4, p. 2344, 2023

    2023

  5. [5]

    KairosAD: A SAM-Based Model for Industrial Anomaly Detection on Embedded Devices,

    U. Khanet al., “KairosAD: A SAM-Based Model for Industrial Anomaly Detection on Embedded Devices,” in23rd International Con- ference on Image Analysis and Processing (ICIAP), 2025

    2025

  6. [6]

    PaSTe: Improving the Efficiency of Visual Anomaly Detection at the Edge,

    M. Baruscoet al., “PaSTe: Improving the Efficiency of Visual Anomaly Detection at the Edge,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2025

    2025

  7. [7]

    Reviewing progresses on In-Sensor AI Computing,

    D. P. Pauet al., “Reviewing progresses on In-Sensor AI Computing,” Microprocessors and Microsystems, p. 105156, 2025

    2025

  8. [8]

    Performance Analysis of Edge and In-Sensor AI Processors: A Comparative Review,

    L. Capogrossoet al., “Performance Analysis of Edge and In-Sensor AI Processors: A Comparative Review,”arXiv preprint arXiv:2603.08725, 2026

    arXiv 2026

Show all 49 references
  1. [9]

    Wide Residual Networks,

    S. Zagoruykoet al., “Wide Residual Networks,” inBritish Machine Vision Conference (BMVC), 2016

    2016

  2. [10]

    Deep Residual Learning for Image Recognition,

    K. Heet al., “Deep Residual Learning for Image Recognition,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  3. [11]

    Model Compression Toolkit (MCT),

    Sony Semiconductor Solutions, “Model Compression Toolkit (MCT),” https://github.com/SonySemiconductorSolutions/ mct-model-optimization, accessed: 2026-03-05

    2026

  4. [12]

    MVTec AD – A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection,

    P. Bergmannet al., “MVTec AD – A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  5. [13]

    Sony Semiconductor Solutions, “IMX500,” https://developer.sony.com/ imx500, accessed: 2025-11-30

    2025

  6. [14]

    Memorizing Normality to Detect Anomaly: Memory- Augmented Deep Autoencoder for Unsupervised Anomaly Detection,

    D. Gonget al., “Memorizing Normality to Detect Anomaly: Memory- Augmented Deep Autoencoder for Unsupervised Anomaly Detection,” inIEEE/CVF International Conference on Computer Vision (ICCV), 2019

    2019

  7. [15]

    Reconstruction by inpainting for visual anomaly detection,

    V . Zavrtaniket al., “Reconstruction by inpainting for visual anomaly detection,”Pattern Recognition, vol. 112, p. 107706, 2021

    2021

  8. [16]

    Self-Supervised Predictive Convolutional Attentive Block for Anomaly Detection,

    N.-C. Risteaet al., “Self-Supervised Predictive Convolutional Attentive Block for Anomaly Detection,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  9. [17]

    Towards Total Recall in Industrial Anomaly Detection,

    K. Rothet al., “Towards Total Recall in Industrial Anomaly Detection,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  10. [18]

    Pni : Industrial anomaly detection using position and neighborhood information,

    J. Baeet al., “Pni : Industrial anomaly detection using position and neighborhood information,” inIEEE/CVF International Conference on Computer Vision (ICCV), 2023

    2023

  11. [19]

    PANDA: Adapting Pretrained Features for Anomaly Detection and Segmentation,

    T. Reisset al., “PANDA: Adapting Pretrained Features for Anomaly Detection and Segmentation,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  12. [20]

    PaDiM: A Patch Distribution Modeling Framework for Anomaly Detection and Localization,

    T. Defardet al., “PaDiM: A Patch Distribution Modeling Framework for Anomaly Detection and Localization,” in25th International Conference on Pattern Recognition Workshops (ICPRW), 2021

    2021

  13. [21]

    FastFlow: Unsupervised Anomaly Detection and Local- ization via 2D Normalizing Flows,

    J. Yuet al., “FastFlow: Unsupervised Anomaly Detection and Local- ization via 2D Normalizing Flows,”arXiv preprint arXiv:2111.07677, 2021

    2021 arXiv

  14. [22]

    Same Same but DifferNet: Semi-Supervised Defect Detection With Normalizing Flows,

    M. Rudolphet al., “Same Same but DifferNet: Semi-Supervised Defect Detection With Normalizing Flows,” inIEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021

    2021

  15. [23]

    CFLOW-AD: Real-Time Unsupervised Anomaly Detection with Localization via Conditional Normalizing Flows,

    D. Gudovskiyet al., “CFLOW-AD: Real-Time Unsupervised Anomaly Detection with Localization via Conditional Normalizing Flows,” in IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022

    2022

  16. [24]

    Diffusion-Based Image Generation for In- Distribution Data Augmentation in Surface Defect Detection,

    L. Capogrossoet al., “Diffusion-Based Image Generation for In- Distribution Data Augmentation in Surface Defect Detection,” in19th International Conference on Computer Vision Theory and Applications (VISAPP), 2024

    2024

  17. [25]

    Leveraging Latent Diffusion Models for Training- Free in-Distribution Data Augmentation for Surface Defect Detection,

    F. Girellaet al., “Leveraging Latent Diffusion Models for Training- Free in-Distribution Data Augmentation for Surface Defect Detection,” inInternational Conference on Content-Based Multimedia Indexing (CBMI), 2024

    2024

  18. [26]

    CutPaste: Self-Supervised Learning for Anomaly Detection and Localization,

    C.-L. Liet al., “CutPaste: Self-Supervised Learning for Anomaly Detection and Localization,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  19. [27]

    DRAEM - A Discriminatively Trained Recon- struction Embedding for Surface Anomaly Detection,

    V . Zavrtaniket al., “DRAEM - A Discriminatively Trained Recon- struction Embedding for Surface Anomaly Detection,” inInternational Conference on Computer Vision (ICCV), 2021

    2021

  20. [28]

    SimpleNet: A Simple Network for Image Anomaly Detection and Localization,

    Z. Liuet al., “SimpleNet: A Simple Network for Image Anomaly Detection and Localization,” inConference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  21. [29]

    A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization,

    W. Chenet al., “A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization,” inEuropean Conference on Computer Vision (ECCV), 2024

    2024

  22. [30]

    A Machine Learning-Oriented Survey on Tiny Machine Learning,

    L. Capogrossoet al., “A Machine Learning-Oriented Survey on Tiny Machine Learning,”IEEE Access, vol. 12, pp. 23 406–23 426, 2024

    2024

  23. [31]

    Survey and Comparison of Milliwatts Micro controllers for Tiny Machine Learning at the Edge,

    M. Giordanoet al., “Survey and Comparison of Milliwatts Micro controllers for Tiny Machine Learning at the Edge,” in4th International Conference on Artificial Intelligence Circuits and Systems (AICAS), 2022

    2022

  24. [32]

    TinyTracker: Ultra-Fast and Ultra-Low-Power Edge Vision In-Sensor for Gaze Estimation,

    P. Bonazziet al., “TinyTracker: Ultra-Fast and Ultra-Low-Power Edge Vision In-Sensor for Gaze Estimation,”IEEE Sensors, 2023

    2023

  25. [33]

    MCUNet: Tiny Deep Learning on IoT Devices,

    J. Linet al., “MCUNet: Tiny Deep Learning on IoT Devices,” in Advances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  26. [34]

    TinyissimoYOLO: A Quantized, Low-Memory Footprint, TinyML Object Detection Network for Low Power Micro- controllers,

    J. Moosmannet al., “TinyissimoYOLO: A Quantized, Low-Memory Footprint, TinyML Object Detection Network for Low Power Micro- controllers,” in5th International Conference on Artificial Intelligence Circuits and Systems (AICAS), 2023

    2023

  27. [35]

    Ultra-Efficient On-Device Object Detection on AI-Integrated Smart Glasses with TinyissimoYOLO,

    J. Moosmannet al., “Ultra-Efficient On-Device Object Detection on AI-Integrated Smart Glasses with TinyissimoYOLO,” inEuropean Con- ference on Computer Vision (ECCV) Workshops, 2024

    2024

  28. [36]

    9.6 A 1/2.3inch 12.3Mpixel with On-Chip 4.97TOPS/W CNN Processor Back-Illuminated Stacked CMOS Image Sensor,

    R. Ekiet al., “9.6 A 1/2.3inch 12.3Mpixel with On-Chip 4.97TOPS/W CNN Processor Back-Illuminated Stacked CMOS Image Sensor,” in IEEE International Solid- State Circuits Conference (ISSCC), 2021

    2021

  29. [37]

    PicoSAM2: Low-Latency Segmentation In-Sensor for Edge Vision Applications,

    P. Bonazziet al., “PicoSAM2: Low-Latency Segmentation In-Sensor for Edge Vision Applications,” inIEEE SENSORS, 2025

    2025

  30. [38]

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

    Q. Tonget al., “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

  31. [39]

    Q-Segment: Segmenting Images In-Sensor for Vessel- Based Medical Diagnosis,

    P. Bonazziet al., “Q-Segment: Segmenting Images In-Sensor for Vessel- Based Medical Diagnosis,” inIEEE 6th International Conference on AI Circuits and Systems (AICAS), 2024

    2024

  32. [40]

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

    T. Cuiet al., “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

  33. [41]

    The Raspberry Pi AI Camera,

    Raspberry Pi Documentation, “The Raspberry Pi AI Camera,” https: //www.raspberrypi.com/documentation/accessories/ai-camera.html, ac- cessed: 2026-03-05

    2026

  34. [42]

    S. J. Prince,Understanding Deep Learning. MIT press, 2023

    2023

  35. [43]

    Focal Loss for Dense Object Detection,

    T.-Y . Linet al., “Focal Loss for Dense Object Detection,” inIEEE International Conference on Computer Vision (ICCV), 2017

    2017

  36. [44]

    Decoupled Weight Decay Regularization,

    I. Loshchilovet al., “Decoupled Weight Decay Regularization,” in International Conference on Learning Representations, 2019

    2019

  37. [45]

    ImageNet: A Large-Scale Hierarchical Image Database,

    J. Denget al., “ImageNet: A Large-Scale Hierarchical Image Database,” inIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009

    2009

  38. [46]

    SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and < 0.5 MB model size,

    F. N. Iandolaet al., “SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and < 0.5 MB model size,” inInternational Conference on Learning Representations (ICLR), 2017

    2017

  39. [47]

    Searching for MobileNetV3,

    A. Howardet al., “Searching for MobileNetV3,” inInternational Con- ference on Computer Vision (ICCV), 2019

    2019

  40. [48]

    EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,

    M. Tanet al., “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” inInternational Conference on Machine Learning (ICML), 2019

    2019

  41. [49]

    SPot-the-Difference Self-supervised Pre-training for Anomaly Detection and Segmentation,

    Y . Zouet al., “SPot-the-Difference Self-supervised Pre-training for Anomaly Detection and Segmentation,” inEuropean Conference on Computer Vision (ECCV), 2022

    2022

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.