Pith. sign in

REVIEW 2 major objections 7 minor 64 references

Optimization of DNN-based HSI Segmentation FPGA-based SoC for ADS: A Practical Approach

T0 review · 2 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Iterative pruning and quantization shrink a hyperspectral U-Net to 1% of its parameters while keeping weighted IoU at 88.37.

desk verdict Useful measured end-to-end HSI segmentation deployment on an FPGA SoC; the main caveat is that accuracy preservation is shown on one fold only. read the letter →

arxiv 2507.16556 v1 pith:GEMTKCJY submitted 2025-07-22 cs.CV cs.AIcs.ARcs.LGeess.IV

classification cs.CVcs.AIcs.ARcs.LGeess.IV
keywords hyperspectralimagingsemanticsegmentationU-Netstructuredpruningiterative8-bitquantizationFPGAsystem-on-chipautonomousdriving
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

This paper claims that a hyperspectral image segmentation network for autonomous driving can be compressed hard enough for an FPGA-based system-on-chip without losing accuracy, if the compression is guided by where the network actually spends its computation. Applying iterative structured pruning and 8-bit quantization to a U-Net trained on the HSI-Drive v2.0 dataset shrinks the model to 0.32 million parameters and 8.49 GFLOPs, about 1% of the original parameter count and 24% of the original operations, while weighted IoU stays at 88.37 versus 88.40 for the uncompressed model. The same optimization flow, together with hardware-aware preprocessing and a three-stage pipeline, makes the whole system run at 10.54 frames per second on the target board, with a 2.86x inference speed-up from pruning alone. A sympathetic reader would care because it demonstrates a practical path from a research dataset to a working embedded pipeline, including the preprocessing stage that most studies ignore.

What carries the argument

The load-bearing mechanism is the iterative structured pruning procedure. Static analysis identifies the convolution and transposed-convolution layers as essentially the whole computational cost, so the pruning objective is a target FLOPs reduction; dynamic analysis measures, for each convolutional layer, how class-level IoU degrades as that layer's channels are removed by smallest L1 norm while the rest of the network stays frozen. A constrained binary search then distributes the global pruning ratio into layer-specific ratios, locking any layer whose predicted degradation exceeds the user tolerance, and the surviving scheme is fine-tuned before another iteration begins. The same toolchain also performs INT8 quantization with batch-normalization folding, and the final preprocessing normalization is folded into a depthwise convolution so the accelerator can execute it.

What would settle it

Apply the reported two-iteration layer-specific pruning scheme (overall 0.75) to each of the other four stratified folds of HSI-Drive v2.0 and measure weighted IoU; if the mean falls more than about one point below the 88.40 uncompressed baseline, the single-layer sensitivity curves are miscalibrated for joint pruning.

Watch

Extended reading notes

Core claim

The central claim is that a U-Net for 25-band hyperspectral driving-scene segmentation can be reduced from 31.10 million parameters and 34.87 GFLOPs to 0.32 million parameters and 8.49 GFLOPs with negligible loss of segmentation quality, and that this reduction is achievable by iterating a post-training, structured channel-pruning procedure that combines a static analysis of per-layer FLOPs and parameter counts with a dynamic per-layer sensitivity analysis. Each convolutional layer is pruned one at a time using the smallest-L1-norm channels as the pruning criterion, and a constrained binary search assigns layer-specific pruning ratios that meet a global FLOPs target while keeping weighted IoU above a threshold. The paper reports that after two pruning iterations (overall ratio 0.75) and fine-tuning, weighted IoU is 88.37 versus 88.40 for the uncompressed INT8 model, and that the pruned model outperforms both one-time pruning at the same global ratio and shallower dense U-Nets trained from scratch.

Load-bearing premise

The pruning ratios come from sensitivity curves measured one layer at a time with the rest of the network frozen, and the method assumes those curves still predict accuracy when many layers are pruned together and then fine-tuned.

Editorial extensions

If this is right

  • At an overall pruning ratio of 0.75 achieved in two iterations, the model keeps wIoU within 0.03 points of the uncompressed INT8 U-Net while cutting parameters by about 99% and operations by about 76%.
  • Iterative pruning at a global ratio of 0.8 reaches wIoU 84.43, nearly one point higher than one-time pruning at 0.8 (83.51), and produces roughly half the parameters, because previously locked layers can be pruned again in the second iteration.
  • Preprocessing, not inference, is the bottleneck: it takes about 2.5x the DPU latency in the single-stage design, and splitting it into two threads reduces the longest task by about 15%.
  • Folding the final symmetric normalization into a depthwise convolution moves it onto the accelerator; requantization changes the winning class for only about 2% of pixels, mostly at class boundaries.
  • Choosing the right DPU configuration and the two-iteration pruning together produce an 8.18x throughput improvement from worst-case to best-case configuration.

Reading between the lines

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

  • A testable extension is to apply the same iterative scheme to the other four stratified folds of HSI-Drive v2.0; the paper reports fold 5 only, so the layer-specific ratios may be miscalibrated on other splits.
  • The sensitivity analysis suggests structural conclusions the paper states only implicitly: layers feeding the skip connections are the least prunable in post-training pruning, which favors asymmetric encoder-heavy, decoder-light architectures.
  • Because preprocessing dominates latency even after three-stage pipelining, further gains would likely come from offloading reflectance correction and demosaicing to programmable logic or fusing them into the network, an option the paper identifies but does not implement.
  • The authors' comparison with pre-training pruning indicates that post-training iterative pruning is more robust across random initializations; a natural next experiment would be to combine the two, using pre-training to identify prunable architectures and post-training sensitivity to refine ratios.
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

2 major / 7 minor

Summary. The paper presents a hardware/software co-design methodology for deploying a U-Net-based HSI segmentation pipeline on an AMD-Xilinx KV260 FPGA-based SoC. The contributions span raw-data preprocessing (cropping, reflectance correction, demosaicing, normalization) with careful memory-format management (BSQ-to-BIP), iterative post-training structured pruning guided by static and dynamic sensitivity analyses, post-training quantization, and a three-stage pipelined deployment. The authors report reducing the baseline U-Net from 31.10 M parameters and 34.87 GFLOPS to 0.32 M parameters and 8.49 GFLOPs (1.02% and 24.34% of the original, respectively) while maintaining a weighted IoU of 88.37 versus 88.40 for the FP32 baseline, and a 2.86x inference speed-up on the B4096 DPU configuration.

Significance. If the claims hold, this is a valuable practical contribution to embedded HSI segmentation for autonomous driving: it demonstrates that aggressive compression (structured pruning plus INT8 quantization) can be combined with a full preprocessing pipeline on a commercially available SoC, and it provides a reproducible algorithm (Algorithm 1) for iterative pruning. The paper is commendable for reporting internally consistent numbers (the abstract matches Table 3), for comparing against a state-of-the-art RGB model (Intern Image), and for including detailed deployment data (power, resource usage, throughput). The main limitation is the use of a single test fold both for pruning-ratio selection and for final evaluation, which weakens the generality of the accuracy-preservation claim.

major comments (2)
  1. [3.2, 4.2.3, Table 3] The central accuracy claim is not an independent evaluation because the same fold used for selecting pruning decisions is also used for reporting the final accuracy. Section 3.2 states that 'from now on... the results provided will correspond to the fifth fold'; Section 4.2.2 builds sensitivity curves from wIoU measured on that fold (Figures 4 and 5); Section 4.2.3 uses Algorithm 1 with thresholds to select layer-specific pruning ratios; and Table 3 rows are compared on the same fold to choose the 0.75 schedule. Therefore, the reported wIoU of 88.37 is a selection result rather than a hold-out result. For an ADS-oriented claim, the authors should evaluate the final compressed model on the remaining folds (or at least report per-fold results with the same pruning ratios) to substantiate 'no noticeable degradation'.
  2. [4.2.2-4.2.3] The sensitivity analysis prunes each layer independently while all other layers are frozen, and the constrained binary search in Algorithm 1 uses these independent curves to assign simultaneous layer-specific pruning ratios. This ignores possible interactions between concurrently pruned layers. The paper partially mitigates this concern by recomputing sensitivity on the pruned model (Figure 5b) and by comparing iterative versus one-time pruning at a global 0.8 ratio (Section 4.2.4). However, the final 0.75 schedule is still derived from independent-sensitivity data on the same fold. Providing at least one validation where the selected ratios are applied jointly and the wIoU is measured on a fold not used for selection (or comparing against a joint-sensitivity estimate) would substantially strengthen the method.
minor comments (7)
  1. [6.4] Section 6.4 contains a typo: 'his created three concurrent threads' should be 'This created three concurrent threads'.
  2. [4.2.4] Section 4.2.4 contains garbled notation: 'overallprof 0.99' and 'The prof the central area' should read 'overall pr' and 'The pr in the central area', respectively.
  3. [4.2.3] Section 4.2.3 has a typo: 'For an initialprof 0.5' should be 'For an initial pr of 0.5'.
  4. [4.2.1, Eq. (1)] Equation (1) includes the factor (1/4)^j on every term, but the text says this factor applies only to transposed conv2d layers. The equation should clarify that the factor is 1 for regular conv layers.
  5. [Table 3] The row labeled '0b' in Table 3 is ambiguous; the footnote says it corresponds to the depth-4 floating-point model from [53], but the label '0' could be mistaken for a pruning ratio of 0. Use a clearer label such as 'FP32 depth-4 [53]'.
  6. [1, 7] The paper claims that the optimization reduces inference operations by 'an order of magnitude' (Introduction and Conclusions), but the reported value is 24.34% of the original, i.e., a 4.1x reduction. Please revise the wording to 'roughly 4x' or clarify the basis of the order-of-magnitude claim.
  7. [4.2.3] The definition of 'locked layers' is confusing: layers with a pruning ratio of 0.9 are described as 'locked', yet the text later suggests they 'could have been further pruned'. Please rephrase to distinguish layers that are excluded from pruning from layers already pruned at the maximum ratio.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: accuracy and speedup are measured outcomes, not derived from pruning targets; the fold-5 selection caveat is a statistical validity issue rather than a circular derivation.

full rationale

The paper's compression claim is an engineering report rather than a first-principles derivation. The reductions in parameters (to 1.02%) and operations (to 24.34%) are direct consequences of the user-set pruning target of 0.75 and the FLOPs-based pruning objective in Eq. 1, but the paper does not present these reductions as a prediction; it presents them as the achieved configuration. The load-bearing accuracy claim (wIoU 88.37 vs. baseline 88.40) and the 2.86x speedup are measured on the deployed model, not computed from the pruning ratios, so they are not equivalent to the inputs by construction. Self-citations to [17,18] for the public HSI-Drive v2.0 dataset, to [52] for the baseline architecture/training, and to [53] for quantization are independently published, and the paper's own Table 2 re-measures the quantization effect, so these citations are not load-bearing circularity. The one substantive caveat is that pruning ratios, including the second iteration ratio of 0.5, are selected using fold-5 wIoU (Sections 4.2.2-4.2.3, Algorithm 1), while the headline results are also reported for the fifth fold (Section 3.2). This means the 'no noticeable degradation' figure is not cross-fold independent and should be read as a selection-influenced result; however, that is a test-set selection and generalization weakness, not an input-output equivalence, so it does not constitute circularity under the stated rules.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claims rest on a handful of user-chosen thresholds and tooling assumptions, but no new physical or mathematical entities are postulated. The pruning targets are hand-selected ratios, and the accuracy is measured rather than derived, so the ledger is dominated by engineering choices rather than fitted science.

free parameters (7)
  • Overall pruning ratio pr (first iteration) = 0.5
    Hand-selected target FLOPS reduction in Algorithm 1; affects which layers are pruned and the final accuracy.
  • Overall pruning ratio pr (second iteration) = 0.5
    Chosen after observing that the first iteration exceeded baseline wIoU, leading to overall 0.75.
  • wIoU tolerance threshold (ΔwIoU < 1) = 1 IoU point
    Stopping criterion in Algorithm 1, Part 2; hand-tuned for this case study.
  • Layer wIoU tolerance (layer ΔwIoU < 0.25) = 0.25 IoU points
    Threshold governing which layers are locked in later iterations (Algorithm 1, Part 3).
  • Locked layers fraction threshold = 25%
    If more than 25% of layers are locked, the scheme is rejected (Algorithm 1, Part 3).
  • Finetuning epochs = 60
    Used after each pruning iteration with learning rate 1e-6; chosen without a reported sweep.
  • Finetuning learning rate = 1e-6
    Fixed after pruning; not justified by a sensitivity study.
assumptions (5)
  • domain assumption The sensitivity analysis prunes each layer independently on a frozen model, and these per-layer curves predict the behavior of the jointly pruned, fine-tuned model.
    Load-bearing for the pruning allocation (Section 4.2.2 and Figure 4). If interactions matter, the layer-specific ratios from the binary search could be suboptimal.
  • domain assumption Equations (1) and (2) approximate total FLOPS and parameters by ignoring non-convolutional layers, stated to be 0.21% and 0.08% of the total.
    This approximation underpins the static analysis in Section 4.2.1 and the reported GFLOPs reductions.
  • domain assumption The VAI Optimizer 3.5 pruning tool implements channel pruning correctly and its FLOPS accounting matches the DPU workload.
    All pruning is performed through this tool [59]; no independent validation of the tool's pruning decisions is provided.
  • domain assumption The fifth test fold is representative of the dataset, so conclusions from pruning on fold 5 generalize.
    Section 3.3 states results will use fold 5 'without loss of generalization'; this is not demonstrated for the compressed models.
  • domain assumption The on-board preprocessing pipeline reproduces the training-time preprocessing closely enough that accuracy on HSI-Drive v2.0 images transfers to the DPU.
    Section 6.2 validates only the normalization requantization (about 2% pixels change class), not the full reflectance correction and demosaicing chain on the board.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimization of DNN-based HSI Segmentation FPGA-based SoC for ADS: A Practical Approach." pith.science (2026). https://pith.science/paper/GEMTKCJY

@misc{pith2026250716556,
  author       = {Pith},
  title        = {Pith review of: Optimization of DNN-based HSI Segmentation FPGA-based SoC for ADS: A Practical Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GEMTKCJY}},
  note         = {Machine review of arXiv:2507.16556}
}
read the original abstract

The use of HSI for autonomous navigation is a promising research field aimed at improving the accuracy and robustness of detection, tracking, and scene understanding systems based on vision sensors. Combining advanced computer algorithms, such as DNNs, with small-size snapshot HSI cameras enhances the reliability of these systems. HSI overcomes intrinsic limitations of greyscale and RGB imaging in depicting physical properties of targets, particularly regarding spectral reflectance and metamerism. Despite promising results in HSI-based vision developments, safety-critical systems like ADS demand strict constraints on latency, resource consumption, and security, motivating the shift of ML workloads to edge platforms. This involves a thorough software/hardware co-design scheme to distribute and optimize the tasks efficiently among the limited resources of computing platforms. With respect to inference, the over-parameterized nature of DNNs poses significant computational challenges for real-time on-the-edge deployment. In addition, the intensive data preprocessing required by HSI, which is frequently overlooked, must be carefully managed in terms of memory arrangement and inter-task communication to enable an efficient integrated pipeline design on a SoC. This work presents a set of optimization techniques for the practical co-design of a DNN-based HSI segmentation processor deployed on a FPGA-based SoC targeted at ADS, including key optimizations such as functional software/hardware task distribution, hardware-aware preprocessing, ML model compression, and a complete pipelined deployment. Applied compression techniques significantly reduce the complexity of the designed DNN to 24.34% of the original operations and to 1.02% of the original number of parameters, achieving a 2.86x speed-up in the inference task without noticeable degradation of the segmentation accuracy.

Figures

Figures reproduced from arXiv: 2507.16556 by the authors.

Figure 1
Figure 1. Diagram of the DNN-based segmentation pipeline. Left, KV260 board, adapted from [ [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Ground truth (left) and false-RGB (right) of image 721, an example of HSI-Drive v2.0 [17, 18]. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. a shows the number of FLOPS for each 𝑐𝑜𝑛𝑣2𝐷 layer in the model. It can be observed that the decoder branch requires roughly twice as many FLOPS as the encoder branch. Additionally, at each depth level, the second convolution block has an equal number of FLOPS in both the encoder and decoder branches. 0 Decoder 1 2 0 Number of GFLOPS10 9 3 1 Base Depth 2 3 4 5 Encoder (a) Number of FLOPS. 0 Decoder 5 0 10 6 Number of… view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Class IoU-based sensitivity analysis on the U-Net [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Weighted IoU-based sensitivity analyses of different U-Net [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Number of parameters (left), operations (centre) and size (right) comparison among Intern Image (blue [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Comparison of the final pr of: 0.5 (blue), 0.5 & 0.5 (red), 0.5 & 0.6 (yellow) and 0.8 (purple) approaches. This trend also holds when applying a second pr of 0.5 (red line in [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Comparison of the pruning schemes of the pre-training approach (black, blue and light brown) and [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Comparison of the pruning schemes of the pre-training approach (black, blue and light brown) and [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Comparison of the pruning schemes of the pre-training approach (black, blue and light brown) and [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Pipeline scheme comprising from raw image preprocessing to DNN inference. In each step, information [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Comparison of cube data organization in memory (left). The hyperspectral Fabry-Perot filter array [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: DPU output when the normalization is performed by software/explicitly. [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: DPU output when the normalization is performed by hardware/implicitly. [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]
Figure 15
Figure 15. Figure 15: Assessment of the impact of pruning iterations and DPU choice on inference throughput (left) and [PITH_FULL_IMAGE:figures/full_fig_p021_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 34 canonical work pages

  1. [1]

    Fully Convolutional Networks for Semantic Segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully Convolutional Networks for Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3431–3440, 2015. doi: 10.1109/CVPR.2015.7298965

  2. [2]

    Review the State-of-the-art Technologies of Semantic Segmentation Based on Deep Learning.Neurocomputing, 493:626–646, 2022

    Yujian Mo, Yan Wu, Xinneng Yang, Feilin Liu, and Yujun Liao. Review the State-of-the-art Technologies of Semantic Segmentation Based on Deep Learning.Neurocomputing, 493:626–646, 2022. doi: 10.1016/j.neucom.2022.01.005

  3. [3]

    Deep Learning in Medical Hyperspectral Images: A Review.Sensors, 22(24):9790, 2022

    Rong Cui, He Yu, Tingfa Xu, Xiaoxue Xing, Xiaorui Cao, Kang Yan, and Jiexi Chen. Deep Learning in Medical Hyperspectral Images: A Review.Sensors, 22(24):9790, 2022. doi: 10.3390/s22249790

  4. [4]

    Current State of Hyperspectral Remote Sensing for Early Plant Disease Detection: A Review.Sensors, 22(3):757, 2022

    Anton Terentev, Viktor Dolzhenko, Alexander Fedotov, and Danila Eremenko. Current State of Hyperspectral Remote Sensing for Early Plant Disease Detection: A Review.Sensors, 22(3):757, 2022. doi: 10.3390/s22030757

  5. [5]

    Rapid and Noninvasive Sensory Analyses of Food Products by Hyperspectral Imaging: Recent Application Developments.Trends in Food Science & Technology, 111:151–165, 2021

    Gözde Özdoğan, Xiaohui Lin, and Da-Wen Sun. Rapid and Noninvasive Sensory Analyses of Food Products by Hyperspectral Imaging: Recent Application Developments.Trends in Food Science & Technology, 111:151–165, 2021. doi: 10.1016/j.tifs.2021.02.044

  6. [6]

    Frequency of Metamerism in Natural Scenes.Journal of the Optical Society of America A, 23(10):2359–2372, 2006

    David H Foster, Kinjiro Amano, Sérgio MC Nascimento, and Michael J Foster. Frequency of Metamerism in Natural Scenes.Journal of the Optical Society of America A, 23(10):2359–2372, 2006. doi: 10.1364/JOSAA.23.002359

  7. [7]

    Victoria Martínez, Unai Martinez-Corral, Óscar Mata- Carballeira, and Inés del Campo

    Jon Gutiérrez-Zaballa, Koldo Basterretxea, Javier Echanobe, M. Victoria Martínez, Unai Martinez-Corral, Óscar Mata- Carballeira, and Inés del Campo. On-chip Hyperspectral Image Segmentation With Fully Convolutional Networks For Scene Understanding in Autonomous Driving.Journal of Systems Architecture, 139:102878, 2023. ISSN 1383-7621. doi: 10.1016/j.sysar...

  8. [8]

    Hyper-Drive: Visible-Short Wave Infrared Hyperspectral Imaging Datasets for Robots in Unstructured Environments

    Nathaniel Hanson, Benjamin Pyatski, Samuel Hibbard, Charles DiMarzio, and Taşkın Padır. Hyper-Drive: Visible-Short Wave Infrared Hyperspectral Imaging Datasets for Robots in Unstructured Environments. In2023 13th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS), pages 1–5. IEEE, 2023. doi: 10.1109/WHISPERS614...

Show all 64 references
  1. [9]

    Point-Supervised Semantic Segmentation of Natural Scenes via Hyperspectral Imaging

    Tianqi Ren, Qiu Shen, Ying Fu, and Shaodi You. Point-Supervised Semantic Segmentation of Natural Scenes via Hyperspectral Imaging. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1357–1367, 2024. doi: 10.1109/CVPRW63...

  2. [10]

    A Tiny VIS-NIR Snapshot Multi- spectral Camera

    Bert Geelen, Carolina Blanch, Pilar Gonzalez, Nicolaas Tack, and Andy Lambrechts. A Tiny VIS-NIR Snapshot Multi- spectral Camera. In Georg von Freymann, Winston V. Schoenfeld, Raymond C. Rumpf, and Henry Helvajian, editors, Advanced Fabrication Technologies for Micro/Nano Opti...

  3. [11]

    Accurate Video-Rate Multi-Spectral Imaging Using IMEC Snapshot Sensors

    Kathleen Vunckx and Wouter Charle. Accurate Video-Rate Multi-Spectral Imaging Using IMEC Snapshot Sensors. In 2021 11th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS), pages 1–7, 2021. doi: 10.1109/WHISPERS52202.2021.9483975

  4. [12]

    A Compact Snapshot Multispectral Imager with A Monolithically Integrated Per-pixel Filter Mosaic

    Bert Geelen, Nicolaas Tack, and Andy Lambrechts. A Compact Snapshot Multispectral Imager with A Monolithically Integrated Per-pixel Filter Mosaic. In Georg von Freymann, Winston V. Schoenfeld, and Raymond C. Rumpf, editors, Advanced Fabrication Technologies for Micro/Nano Opti...

  5. [13]

    Kria K26 SOM: The Ideal Platform for Vision AI at the Edge

    AMD-Xilinx. Kria K26 SOM: The Ideal Platform for Vision AI at the Edge. https://docs.xilinx.com/v/u/en-US/wp529- som-benchmarks, 2021

  6. [14]

    Kria K26 SOM Data Sheet (DS987)

    AMD-Xilinx. Kria K26 SOM Data Sheet (DS987). https://docs.amd.com/r/en-US/ds987-k26-som/Overview, 2024

  7. [15]

    Arm Cortex-A53 MPCore Processor Technical Reference Manual

    ARM. Arm Cortex-A53 MPCore Processor Technical Reference Manual. https://developer.arm.com/documentation/ ddi0500, 2018. Accessed: 2024-11-08

  8. [16]

    DPUCZDX8G for Zynq UltraScale+ MPSoCs Product Guide (PG338)

    Xilinx. DPUCZDX8G for Zynq UltraScale+ MPSoCs Product Guide (PG338). https://https://docs.xilinx.com/r/en- US/pg338-dpu?tocId=3xsG16y_QFTWvAJKHbisEw, 2024

  9. [17]

    HSI-Drive v2.0: More Data for New Challenges in Scene Understanding for Autonomous Driving

    Jon Gutiérrez-Zaballa, Koldo Basterretxea, Javier Echanobe, M Victoria Martínez, and Unai Martinez-Corral. HSI-Drive v2.0: More Data for New Challenges in Scene Understanding for Autonomous Driving. In2023 IEEE Symposium Series on Computational Intelligence (SSCI), pages 207–2...

  10. [18]

    Hsi -drive, 2023

    Koldo Basterretxea, Jon Gutiérrez -Zaballa, Javier Echanobe, and María Victoria Martínez. Hsi -drive, 2023. URL https://doi.org/10.5281/zenodo.15686957

  11. [19]

    U-Net: Convolutional Networks for Biomedical Image Segmenta- tion

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmenta- tion. InMedical Image Computing and Computer-assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part...

  12. [20]

    HyKo: A Spectral Dataset for Scene Understanding

    Christian Winkens, Florian Sattler, Veronika Adams, and Dietrich Paulus. HyKo: A Spectral Dataset for Scene Understanding. InProceedings of the IEEE International Conference on Computer Vision Workshops (ICCVW), pages 254–261, 2017. doi: 10.1109/ICCVW.2017.39

  13. [21]

    HS3-Bench: A Benchmark and Strong Baseline for Hyperspectral Semantic Segmentation in Driving Scenarios

    Nick Theisen, Robin Bartsch, Dietrich Paulus, and Peer Neubert. HS3-Bench: A Benchmark and Strong Baseline for Hyperspectral Semantic Segmentation in Driving Scenarios. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5895–5901. IEEE, 20...

  14. [22]

    Hyperspectral City V1.0 Dataset and Benchmark.arXiv, 2019

    Shaodi You, Erqi Huang, Shuaizhe Liang, Yongrong Zheng, Yunxiang Li, Fan Wang, Sen Lin, Qiu Shen, Xun Cao, Diming Zhang, et al. Hyperspectral City V1.0 Dataset and Benchmark.arXiv, 2019. doi: 10.48550/arXiv.1907.10270

  15. [23]

    HSICityV2: Urban Scene Understanding via Hyper- spectral Images, Jul 2021

    Yuxing Huang, Tianqi Ren, Qiu Shen, Ying Fu, and Shaodi You. HSICityV2: Urban Scene Understanding via Hyper- spectral Images, Jul 2021

  16. [24]

    Hsi Road: A Hyper Spectral Image Dataset for Road Segmentation

    Jiarou Lu, Huafeng Liu, Yazhou Yao, Shuyin Tao, Zhenming Tang, and Jianfeng Lu. Hsi Road: A Hyper Spectral Image Dataset for Road Segmentation. In2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6. IEEE,

  17. [25]

    Basterretxea, V

    K. Basterretxea, V. Martínez, J. Echanobe, J. Gutiérrez–Zaballa, and I. Del Campo. HSI-Drive: A Dataset for the Research of Hyperspectral Image Processing Applied to Autonomous Driving Systems. In2021 IEEE Intelligent Vehicles Symposium (IV), pages 866–873, 2021. doi: 10.1109/...

  18. [26]

    The Cityscapes Dataset for Semantic Urban Scene Understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes Dataset for Semantic Urban Scene Understanding. InProceedings of the IEEE Conference on Computer Vision and Pattern Recog...

  19. [28]

    Semantic Object Classes in Video: A High-definition Ground Truth Database.Pattern Recognition Letters, 30(2):88–97, 2009

    Gabriel J Brostow, Julien Fauqueur, and Roberto Cipolla. Semantic Object Classes in Video: A High-definition Ground Truth Database.Pattern Recognition Letters, 30(2):88–97, 2009. doi: 10.1016/j.patrec.2008.04.005

  20. [29]

    The Mapillary Vistas Dataset for Semantic Understanding of Street Scenes

    Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The Mapillary Vistas Dataset for Semantic Understanding of Street Scenes. InProceedings of the IEEE International Conference on Computer Vision (ICCV), pages 4990–4999, 2017. doi: 10.1109/ICCV.2017.534

  21. [30]

    The ApolloScape Dataset for Autonomous Driving

    Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The ApolloScape Dataset for Autonomous Driving. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 954–960, 2018. doi: 1...

  22. [32]

    Cityscapes Dataset Benchmarks, 2024

    Cityscapes Dataset. Cityscapes Dataset Benchmarks, 2024. URL https://www.cityscapes-dataset.com/benchmarks/. Accessed: September 16, 2024

  23. [33]

    Strong but simple: A Baseline for Domain Generalized Dense Perception by CLIP-based Transfer Learning.arXiv, 2023

    Christoph Hümmer, Manuel Schwonberg, Liangwei Zhong, Hu Cao, Alois Knoll, and Hanno Gottschalk. Strong but simple: A Baseline for Domain Generalized Dense Perception by CLIP-based Transfer Learning.arXiv, 2023. doi: 10.48550/arXiv.2312.02021

  24. [34]

    To Prune, or Not to Prune: Exploring the Efficacy of Pruning for Model Compression

    Michael Zhu and Suyog Gupta. To Prune, or Not to Prune: Exploring the Efficacy of Pruning for Model Compression. arXiv, 2017. doi: 10.48550/arXiv.1710.01878. , Vol. 1, No. 1, Article . Publication date: September 2025. 26 Gutiérrez-Zaballa et al

  25. [35]

    A Dynamic CNN Pruning Method Based on Matrix Similarity.Signal, Image and Video Processing, 15:381–389, 2021

    Mingwen Shao, Junhui Dai, Jiandong Kuang, and Deyu Meng. A Dynamic CNN Pruning Method Based on Matrix Similarity.Signal, Image and Video Processing, 15:381–389, 2021. doi: 10.1007/s11760-020-01760-x

  26. [36]

    Importance Estimation for Neural Network Pruning

    Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance Estimation for Neural Network Pruning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11264– 11272, 2019. doi: 10.1109/CVPR.2019.01152

  27. [37]

    Optimal Brain Compression: A Framework for Accurate Post-Training Quantization and Pruning

    Elias Frantar and Dan Alistarh. Optimal Brain Compression: A Framework for Accurate Post-Training Quantization and Pruning. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems, volume 35, pages 4475–44...

  28. [38]

    A Fast Post- Training Pruning Framework for Transformers

    Woosuk Kwon, Sehoon Kim, Michael W Mahoney, Joseph Hassoun, Kurt Keutzer, and Amir Gholami. A Fast Post- Training Pruning Framework for Transformers. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems...

  29. [39]

    Plug-and-play: An Efficient Post-training Pruning Method for Large Language Models

    Yingtao Zhang, Haoli Bai, Haokun Lin, Jialin Zhao, Lu Hou, and Carlo Vittorio Cannistraci. Plug-and-play: An Efficient Post-training Pruning Method for Large Language Models. InThe Twelfth International Conference on Learning Representations, 2024

  30. [40]

    The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks

    Jonathan Frankle and Michael Carbin. The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks. arXiv, 2018. doi: 10.48550/arXiv.1803.03635

  31. [41]

    Rethinking the Value of Network Pruning

    Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the Value of Network Pruning. arXiv, 2018. doi: 10.48550/arXiv.1810.05270

  32. [42]

    Deconstructing Lottery Tickets: Zeros, Signs, and the Supermask.Advances in Neural Information Processing Systems, 32, 2019

    Hattie Zhou, Janice Lan, Rosanne Liu, and Jason Yosinski. Deconstructing Lottery Tickets: Zeros, Signs, and the Supermask.Advances in Neural Information Processing Systems, 32, 2019

  33. [43]

    What’s Hidden in A Randomly Weighted Neural Network? InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11893–11902, 2020

    Vivek Ramanujan, Mitchell Wortsman, Aniruddha Kembhavi, Ali Farhadi, and Mohammad Rastegari. What’s Hidden in A Randomly Weighted Neural Network? InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11893–11902, 2020. doi: 10.1109/C...

  34. [44]

    Proving the Lottery Ticket Hypothesis: Pruning is All You Need

    Eran Malach, Gilad Yehudai, Shai Shalev-Schwartz, and Ohad Shamir. Proving the Lottery Ticket Hypothesis: Pruning is All You Need. InInternational Conference on Machine Learning, pages 6682–6691. PMLR, 2020

  35. [45]

    A Comprehensive Review of Network Pruning Based on Pruning Granularity and Pruning Time Perspectives.Neurocomputing, page 129382, 2025

    Kehan Zhu, Fuyi Hu, Yuanbing Ding, Wei Zhou, and Ruxin Wang. A Comprehensive Review of Network Pruning Based on Pruning Granularity and Pruning Time Perspectives.Neurocomputing, page 129382, 2025. doi: 10.1016/j. neucom.2025.129382

  36. [46]

    A Survey on Deep Neural Network Pruning: Taxonomy, Comparison, Analysis, and Recommendations.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    Hongrong Cheng, Miao Zhang, and Javen Qinfeng Shi. A Survey on Deep Neural Network Pruning: Taxonomy, Comparison, Analysis, and Recommendations.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. doi: 10.1109/TPAMI.2024.3447085

  37. [47]

    Auto-compressing Subset Pruning for Semantic Image Segmentation

    Konstantin Ditschuneit and Johannes S Otterbach. Auto-compressing Subset Pruning for Semantic Image Segmentation. InDAGM German Conference on Pattern Recognition, pages 20–35. Springer, 2022. doi: 10.1007/978-3-031-16788-1_2

  38. [48]

    Finding Lottery Tickets in Vision Models via Data-driven Spectral Foresight Pruning

    Leonardo Iurada, Marco Ciccone, and Tatiana Tommasi. Finding Lottery Tickets in Vision Models via Data-driven Spectral Foresight Pruning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16142–16151, 2024. doi: 10.1109/CVPR52733...

  39. [49]

    STAMP: Simultaneous Training and Model Pruning for Low Data Regimes in Medical Image Segmentation.Medical Image Analysis, 81:102583, 2022

    Nicola K Dinsdale, Mark Jenkinson, and Ana IL Namburete. STAMP: Simultaneous Training and Model Pruning for Low Data Regimes in Medical Image Segmentation.Medical Image Analysis, 81:102583, 2022. doi: 10.1016/j.media.2022. 102583

  40. [50]

    Dynamically Pruning Segformer for Efficient Semantic Segmentation

    Haoli Bai, Hongda Mao, and Dinesh Nair. Dynamically Pruning Segformer for Efficient Semantic Segmentation. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3298–3302. IEEE, 2022. doi: 10.1109/ICASSP43922.2022.9747634

  41. [51]

    Pruning Parameterization with Bi-level Optimization for Efficient Semantic Segmentation on the Edge

    Changdi Yang, Pu Zhao, Yanyu Li, Wei Niu, Jiexiong Guan, Hao Tang, Minghai Qin, Bin Ren, Xue Lin, and Yanzhi Wang. Pruning Parameterization with Bi-level Optimization for Efficient Semantic Segmentation on the Edge. In Proceedings of the IEEE/CVF Conference on Computer Vision ...

  42. [52]

    Evaluating Single Event Upsets in Deep Neural Networks for Semantic Segmentation: An Embedded System Perspective.Journal of Systems Architecture, 154:103242,

    Jon Gutiérrez-Zaballa, Koldo Basterretxea, and Javier Echanobe. Evaluating Single Event Upsets in Deep Neural Networks for Semantic Segmentation: An Embedded System Perspective.Journal of Systems Architecture, 154:103242,

  43. [53]

    Victoria Martínez

    Jon Gutiérrez-Zaballa, Koldo Basterretxea, Javier Echanobe, Óscar Mata-Carballeira, and M. Victoria Martínez. Rapid Deployment of Domain-specific Hyperspectral Image Processors with Application to Autonomous Driving. In2023 30th IEEE International Conference on Electronics, Ci...

  44. [54]

    Learn the Architecture: Introducing Neon

    ARM. Learn the Architecture: Introducing Neon. https://developer.arm.com/documentation/102474/0100/?lang=en,

  45. [55]

    Deep Learning with INT8 Optimization on Xilinx Devices

    Yao Fu, Ephrem Wu, Ashish Sirasao, Sedny Attia, Kamran Khan, and Ralph Wittig. Deep Learning with INT8 Optimization on Xilinx Devices. https://docs.xilinx.com/v/u/en-US/wp486-deep-learning-int8, 2016

  46. [56]

    Shift: A Zero Flop, Zero Parameter Alternative to Spatial Convolutions

    Bichen Wu, Alvin Wan, Xiangyu Yue, Peter Jin, Sicheng Zhao, Noah Golmant, Amir Gholaminejad, Joseph Gonzalez, and Kurt Keutzer. Shift: A Zero Flop, Zero Parameter Alternative to Spatial Convolutions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog...

  47. [57]

    Accessed: 2024-11-08. , Vol. 1, No. 1, Article . Publication date: September 2025. Optimization of DNN-based HSI Segmentation FPGA-based SoC for ADS: A Practical Approach 27

  48. [58]

    Pruning Filters for Efficient ConvNets, 2017

    Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning Filters for Efficient ConvNets, 2017

  49. [59]

    Vitis AI User Guide

    AMD-Xilinx. Vitis AI User Guide. UG1414 (v3.5) September 28, 2023. https://docs.xilinx.com/r/en-US/ug1414-vitis- ai/Vitis-AI-Overview, 2023

  50. [60]

    Optimal Brain Damage

    Yann LeCun, John Denker, and Sara Solla. Optimal Brain Damage. In D. Touretzky, editor,Advances in Neural Information Processing Systems, volume 2. Morgan-Kaufmann, 1989

  51. [61]

    Encoder-decoder with Atrous Separable Convolution for Semantic Image Segmentation

    Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with Atrous Separable Convolution for Semantic Image Segmentation. InProceedings of the European Conference on Computer Vision (ECCV), pages 801–818, 2018. doi: 10.1007/978-3-030...

  52. [62]

    OpenMP: An Industry Standard API for Shared-memory Programming.IEEE Computational Science and Engineering, 5(1):46–55, 1998

    Leonardo Dagum and Ramesh Menon. OpenMP: An Industry Standard API for Shared-memory Programming.IEEE Computational Science and Engineering, 5(1):46–55, 1998. doi: 10.1109/99.660313

  53. [63]

    Real-time Semantic Image Segmentation with Deep Learning for Autonomous Driving: A Survey.Applied Sciences, 11(19):8802, 2021

    Ilias Papadeas, Lazaros Tsochatzidis, Angelos Amanatiadis, and Ioannis Pratikakis. Real-time Semantic Image Segmentation with Deep Learning for Autonomous Driving: A Survey.Applied Sciences, 11(19):8802, 2021. doi: 10.3390/app11198802

  54. [66]

    Addison-Wesley Professional, 1997

    David R Butenhof.Programming with POSIX Threads. Addison-Wesley Professional, 1997. , Vol. 1, No. 1, Article . Publication date: September 2025

  55. [2020]

    doi: 10.1109/ICME46284.2020.9102890

  56. [2024]

    doi: 10.1016/j.sysarc.2024.103242

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.