Pith. sign in

REVIEW 3 major objections 5 minor 38 references

SpikeYOLO SNNs match conventional deep learning on real automotive detection and tracking, establishing neuromorphic viability for vehicle perception.

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

2026-07-11 11:25 UTC pith:FOCIQRAG

load-bearing objection Solid first application of SpikeYOLO to automotive MOT with competitive two-class numbers; energy-efficiency claim is promissory because nothing is measured on neuromorphic hardware. the 3 major comments →

arxiv 2607.04921 v1 pith:FOCIQRAG submitted 2026-07-06 cs.CV cs.AI

Efficient Perception in Automotive Detection and Tracking Using Neuromorphic Computing

Manish Kolachalam , Rani Malhotra This is my paper
classification cs.CV cs.AI
keywords spiking neural networksSpikeYOLOautomotive object detectionmulti-object trackingneuromorphic computingedge AIKITTIBDD100K
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 spiking neural networks—brain-inspired models that fire only when needed—can handle the core perception jobs of a self-driving car: finding cars and pedestrians and keeping track of them across frames. Using transfer learning on a SpikeYOLO architecture, the authors reach mean average precision of 0.937 on KITTI and 0.771 on BDD100K for detection, and Higher Order Tracking Accuracy of 0.701 and 0.445 respectively—numbers they argue sit in the same range as ordinary deep networks. Because spikes are sparse and asynchronous, the same accuracy can in principle run at far lower energy, which matters for edge devices and electric-vehicle range. The work is the first full evaluation of SNNs on real automotive multi-object tracking benchmarks, so a reader who cares about sustainable autonomy now has concrete evidence that the accuracy gap is closing.

Core claim

Transfer learning with SpikeYOLO yields detection mAP 0.937 (KITTI) and 0.771 (BDD100K MOT2020) and tracking HOTA 0.701 / 0.445, results competitive with conventional deep learning and therefore the first demonstration that SNNs are viable for real-world automotive multi-object perception.

What carries the argument

SpikeYOLO: a YOLO-style hierarchy of integer Leaky-Integrate-and-Fire neurons trained with integer-valued activations and surrogate gradients, then converted to pure spikes only at inference; detections are handed to BoT-SORT for temporal association.

Load-bearing premise

That high accuracy measured on ordinary GPUs, together with energy claims from earlier SpikeYOLO work, is enough to prove energy-efficient viability for autonomous systems even though the present study never runs on neuromorphic chips or reports its own power numbers.

What would settle it

Deploy the same SpikeYOLO weights on a neuromorphic chip (e.g., Loihi 2 or Akida) and measure wall-clock energy and latency on the KITTI and BDD100K sequences; if the measured joules-per-frame are not substantially lower than a comparable GPU YOLO while accuracy holds, the energy-efficiency claim fails.

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

If this is right

  • Automotive perception pipelines can begin swapping conventional detectors for SpikeYOLO-class SNNs without an immediate accuracy penalty.
  • Power budgets for multi-camera, multi-sensor vehicles become smaller once the same accuracy is available in sparse spike form.
  • Event-camera + SNN stacks become a practical next step for further latency and energy gains.
  • On-chip unsupervised adaptation (STDP-style) can be explored for continuous domain shift without cloud retraining.

Where Pith is reading between the lines

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

  • If the energy numbers later materialize on hardware, electric-vehicle range or smaller edge compute boards become the first commercial beneficiaries.
  • The two-class simplification (car / pedestrian) leaves open whether the same architecture scales to the full long-tail of traffic objects without accuracy collapse.
  • Hybrid ANN–SNN pipelines on heterogeneous chips are the natural engineering path once both accuracy and energy are verified.

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 / 5 minor

Summary. The manuscript presents the first evaluation of a spiking neural network (SpikeYOLO, transferred from a COCO-pretrained model) for multi-object detection and tracking on automotive benchmarks (KITTI and BDD100K MOT2020). After class consolidation to car and pedestrian, the authors report detection mAP of 0.937 (KITTI) and 0.771 (BDD100K) and tracking HOTA of 0.701 (KITTI) and 0.445 (BDD100K), numbers they argue are competitive with conventional deep-learning baselines (Tables II–VIII). The architecture, training protocol (SGD, mosaic/HSV/flip augmentation, BoT-SORT association), and evaluation metrics are described in detail; energy-efficiency claims rest on the properties of SNNs and on the cited SpikeYOLO paper rather than new hardware measurements.

Significance. If the accuracy numbers hold under the stated protocol, the work supplies a useful empirical data point: a modern SNN detector/tracker can reach near-SOTA two-class performance on standard automotive sequences without architectural invention. That is a legitimate contribution to the still-sparse literature on SNNs for real-world multi-object perception. The energy-efficiency half of the claim, however, is not measured here; the paper therefore cannot yet establish the full title-level conclusion of “efficient perception … using neuromorphic computing.” The value of the work is primarily as a carefully documented transfer-learning baseline rather than as a demonstrated energy or hardware result.

major comments (3)
  1. Abstract, title and Discussion §5 assert that the reported mAP/HOTA numbers demonstrate high-performance detection and tracking “in an energy-efficient manner” and thereby establish SNN viability for real-world autonomous systems. Methods 3.3 and Table I document only GPU training/inference (RTX 4070/5000, AMP, batch size 4); Discussion §5 explicitly states that deployment remains on conventional GPUs and that full benefits require Loihi 2 or AKIDA. No power, latency, synaptic-operation counts or FLOPs-to-spike conversion figures are supplied for the fine-tuned automotive models. The energy half of the central claim therefore rests solely on citation of the COCO-trained SpikeYOLO paper [10]. Either (a) add a concrete energy/latency bridge measurement for the present models or (b) substantially qualify the abstract, title and concluding claims so that they match what is actually measured.
  2. Methods 3.1 and Results: all four-wheeled vehicles are merged into a single “car” class, person classes into “pedestrian,” and cyclists are excluded; BDD100K experiments use only a 10 K-image subset plus two MOT2020 training splits. Table VIII then compares these two-class, subset numbers against published multi-class SOTA figures (e.g., YOLOPv2 traffic-object mAP, ByteTrack mHOTA). The comparison is therefore not apples-to-apples. Either re-evaluate under the original multi-class protocols or clearly label Table VIII as an approximate, protocol-adjusted comparison and avoid claiming direct competitiveness with the cited multi-class benchmarks.
  3. Tracking protocol (Methods 3.3.4–3.3.5, Tables VI–VII): the KITTI tracking evaluation uses the 21 training sequences rather than a held-out test split, and the same detection model is applied without further adaptation. While common for exploratory work, this choice inflates the risk of optimistic HOTA numbers and weakens the claim of “real-world” tracking viability. Report results on the official KITTI tracking test set (or an explicit held-out split) or qualify the tracking claims accordingly.
minor comments (5)
  1. Table II header and text inconsistently write mAP@50:90 / mAP@50:95; standardize to the conventional mAP@50:95 notation used elsewhere.
  2. Fig. 1 caption and §3.2 describe the architecture at a high level but never state the exact number of time steps / simulation steps used at inference; this is needed for any future energy comparison.
  3. References [10] and [34] are arXiv preprints; update with final venue/DOI if available, and ensure all arXiv links are stable.
  4. §3.4.1 states that overall mAP is the average across classes using their respective IoU thresholds; make the exact averaging formula explicit so that Table II/III numbers can be reproduced.
  5. Author biographies appear after the references; move them to the conventional location or remove if the target journal does not use them.

Circularity Check

0 steps flagged

No significant circularity: empirical mAP/HOTA on external public benchmarks (KITTI, BDD100K) via transfer learning from an independent architecture citation; energy claim is promissory but not a by-construction reduction.

full rationale

The paper's load-bearing results are measured detection (mAP 0.937/0.771) and tracking (HOTA 0.701/0.445) numbers obtained by fine-tuning the COCO-pretrained SpikeYOLO model of Luo et al. [10] (independent authors) on the KITTI and BDD100K training splits and scoring against held-out ground truth under standard protocols (Tables II–VII, Methods 3.1–3.4). These quantities are not fitted parameters renamed as predictions, nor are they definitionally equivalent to any input equation. The architecture description (stem/backbone/neck/I-LIF/DFL) is taken from the external citation and then experimentally validated; no uniqueness theorem, ansatz, or self-citation by the present authors is used to force the numbers. Energy-efficiency language in the Abstract and Discussion §5 is inherited from the general SNN literature and from [10] without new power/latency/synaptic-op measurements on the automotive models (explicitly acknowledged: “deployment remains on traditional GPU hardware”), but this is an unmeasured claim rather than a circular derivation. The evaluation is therefore self-contained against public benchmarks and exhibits none of the six enumerated circularity patterns.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard computer-vision and SNN practice plus a handful of modeling choices (class merging, subset selection, transfer from COCO-pretrained SpikeYOLO). No new physical entities are invented; free parameters are ordinary training hyper-parameters and tracker thresholds. The energy-efficiency half of the claim is imported wholesale from the cited SpikeYOLO paper rather than measured here.

free parameters (4)
  • SGD learning rate / momentum / weight decay
    Set to 0.01 / 0.937 / 0.0005 (Table I); standard but chosen by hand and affect final mAP.
  • Detection confidence threshold 0.25 and BoT-SORT thresholds (init 0.6, IoU 0.3, max age 30)
    Directly gate which boxes enter tracking and therefore control reported HOTA (§3.3.4–3.3.5).
  • Number of fine-tuning epochs (70 KITTI; 45+10 BDD100K) and input resolutions
    Chosen by authors; different schedules would change the accuracy numbers that support the claim.
  • Class-merging rule (all 4-wheel vehicles → car; person classes → pedestrian; cyclists dropped)
    Simplifies the task relative to full multi-class KITTI/BDD benchmarks; performance numbers are conditioned on this choice (§3.1).
axioms (4)
  • domain assumption Surrogate-gradient BPTT through I-LIF neurons yields a faithful approximation to true spiking dynamics for detection accuracy.
    Invoked throughout §2.2–2.3 and §3.2; standard in modern SNN training but still an approximation.
  • domain assumption Integer-valued training followed by spike conversion at inference preserves the accuracy measured on GPU.
    Core of SpikeYOLO [10]; assumed without re-measurement on neuromorphic hardware (§3.2.9).
  • domain assumption BoT-SORT with default hyper-parameters is a fair, state-of-the-art tracker for comparing detection quality.
    Used unchanged (§3.3.4); tracking numbers therefore inherit BoT-SORT’s own assumptions.
  • ad hoc to paper mAP and HOTA on the chosen two-class, subset splits are sufficient proxies for real-world automotive perception viability.
    Stated in Abstract and Discussion; the paper never evaluates full multi-class or full BDD100K.

pith-pipeline@v1.1.0-grok45 · 17574 in / 3150 out tokens · 23268 ms · 2026-07-11T11:25:53.046538+00:00 · methodology

0 comments
read the original abstract

Deep learning algorithms are notorious for their high carbon footprint and computational demands that limit their deployment on edge devices and raise concerns about their long-term sustainability. Neuromorphic computing and Spiking Neural Networks (SNNs) offer a promising alternative to traditional Von Neumann architectures, providing energy-efficient performance, massively parallel computation, and on-chip learning capabilities. Autonomous machines represent a critical application domain where these advantages are particularly valuable. We present the first comprehensive evaluation of SNNs for real-world automotive multi-object detection and tracking. Using transfer learning with the SpikeYOLO architecture, we achieve mean Average Precision of 0.937 on the KITTI dataset and 0.771 on BDD100K MOT2020 dataset for object detection and a Higher Order Tracking Accuracy score of 0.701 (KITTI) and 0.445 (BDD100K MOT2020) for object tracking--results competitive with conventional deep learning methods. Our results demonstrate that SNNs can deliver high-performance object detection and tracking in an energy efficient manner, establishing their viability for perception in real-world autonomous systems.

Figures

Figures reproduced from arXiv: 2607.04921 by Manish Kolachalam, Rani Malhotra.

Figure 1
Figure 1. Figure 1: SpikeYOLO architecture with stem, backbone, neck, and multi [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

38 extracted references · 7 canonical work pages

  1. [1]

    neurons that fire together, wire together,

    or the AKIDA brainchip [8] and deploying them on the edge at scale. A particularly relevant line of research has developed SNNs based on the famous YOLO architecture [9], which provides strong performance on multiple computer vision tasks. These networks, christened SpikeYOLO, have shown strong performance in object detection tasks [10]. Here we validated...

    1949

  2. [2]

    Networks of spiking neurons: The third generation of neural network models,

    W. Maass, “Networks of spiking neurons: The third generation of neural network models,” Neural Networks, vol. 10, no. 9, pp. 1659–1671, Dec. 1997, doi: 10.1016/S0893-6080(97)00011-7

    doi:10.1016/s0893-6080(97)00011-7 1997

  3. [3]

    Towards spike-based machine intelligence with neuromorphic computing,

    K. Roy, A. Jaiswal, and P. Panda, “Towards spike-based machine intelligence with neuromorphic computing,” Nature, vol. 575, no. 7784, pp. 607–617, Nov. 2019, doi: 10.1038/s41586-019-1677-2

    doi:10.1038/s41586-019-1677-2 2019

  4. [4]

    Neuromorphic electronic systems,

    C. Mead, “Neuromorphic electronic systems,” Proc. IEEE, vol. 78, no. 10, pp. 1629–1636, Oct. 1990, doi: 10.1109/5.58356

    doi:10.1109/5.58356 1990

  5. [5]

    Memory and Information Processing in Neuromorphic Systems,

    G. Indiveri and S.-C. Liu, “Memory and Information Processing in Neuromorphic Systems,” Proc. IEEE, vol. 103, no. 8, pp. 1379–1397, Aug. 2015, doi: 10.1109/JPROC.2015.2444094

    doi:10.1109/jproc.2015.2444094 2015

  6. [6]

    A Survey of Neuromorphic Computing and Neural Networks in Hardware,

    C. D. Schuman et al., “A Survey of Neuromorphic Computing and Neural Networks in Hardware,” 2017, arXiv. doi: 10.48550/ARXIV.1705.06963

    Pith/arXiv arXiv doi:10.48550/arxiv.1705.06963 2017

  7. [7]

    A quantitative description of membrane current and its application to conduction and excitation in nerve,

    A. L. Hodgkin and A. F. Huxley, “A quantitative description of membrane current and its application to conduction and excitation in nerve,” J Physiol, vol. 117, no. 4, pp. 500–544, Aug. 1952, doi: 10.1113/jphysiol.1952.sp004764

    doi:10.1113/jphysiol.1952.sp004764 1952

  8. [8]

    Loihi: A Neuromorphic Manycore Processor with On-Chip Learning,

    M. Davies et al., “Loihi: A Neuromorphic Manycore Processor with On-Chip Learning,” IEEE Micro, vol. 38, no. 1, pp. 82–99, Jan. 2018, doi: 10.1109/MM.2018.112130359

    doi:10.1109/mm.2018.112130359 2018

  9. [9]

    BrainChip Showcases AI Benchmarks & Improved Edge Device Metrics,

    “BrainChip Showcases AI Benchmarks & Improved Edge Device Metrics,” BrainChip. Accessed: Dec. 07,

  10. [10]

    Available: https://brainchip.com/brainchip-showcases-compelling- benchmarks-and-recommends-better-metrics-for-ai- devices-at-the-edge/

    [Online]. Available: https://brainchip.com/brainchip-showcases-compelling- benchmarks-and-recommends-better-metrics-for-ai- devices-at-the-edge/

  11. [11]

    You Only Look Once: Unified, Real-Time Object Detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You Only Look Once: Unified, Real-Time Object Detection,” May 09, 2016, arXiv: arXiv:1506.02640. doi: 10.48550/arXiv.1506.02640

    Pith/arXiv arXiv doi:10.48550/arxiv.1506.02640 2016

  12. [12]

    Integer- Valued Training and Spike-Driven Inference Spiking Neural Network for High-performance and Energy- efficient Object Detection,

    X. Luo, M. Yao, Y. Chou, B. Xu, and G. Li, “Integer- Valued Training and Spike-Driven Inference Spiking Neural Network for High-performance and Energy- efficient Object Detection,” Apr. 15, 2025, arXiv: arXiv:2407.20708. doi: 10.48550/arXiv.2407.20708

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.20708 2025

  13. [13]

    Lapicque’s introduction of the integrate- and-fire model neuron (1907),

    L. F. Abbott, “Lapicque’s introduction of the integrate- and-fire model neuron (1907),” Brain Res Bull, vol. 50, no. 5–6, pp. 303–304, 1999, doi: 10.1016/s0361- 9230(99)00161-6

    doi:10.1016/s0361- 1907

  14. [14]

    A Review of the Integrate-and-fire Neuron Model: I. Homogeneous Synaptic Input,

    A. N. Burkitt, “A Review of the Integrate-and-fire Neuron Model: I. Homogeneous Synaptic Input,” Biol Cybern, vol. 95, no. 1, pp. 1–19, Jul. 2006, doi: 10.1007/s00422-006-0068-6

    work page doi:10.1007/s00422-006-0068-6 2006

  15. [15]

    A multisynaptic spiking neuron for simultaneously encoding spatiotemporal dynamics,

    L. Fan et al., “A multisynaptic spiking neuron for simultaneously encoding spatiotemporal dynamics,” Nat Commun, vol. 16, no. 1, p. 7155, Aug. 2025, doi: 10.1038/s41467-025-62251-6

    work page doi:10.1038/s41467-025-62251-6 2025

  16. [16]

    Temporal dendritic heterogeneity incorporated with spiking neural networks for learning multi-timescale dynamics,

    H. Zheng et al., “Temporal dendritic heterogeneity incorporated with spiking neural networks for learning multi-timescale dynamics,” Nat Commun, vol. 15, no. 1, p. 277, Jan. 2024, doi: 10.1038/s41467-023-44614-z

    work page doi:10.1038/s41467-023-44614-z 2024

  17. [17]

    Synaptic plasticity: taming the beast,

    L. F. Abbott and S. B. Nelson, “Synaptic plasticity: taming the beast,” Nat Neurosci, vol. 3, no. S11, pp. 1178–1183, Nov. 2000, doi: 10.1038/81453

    doi:10.1038/81453 2000

  18. [18]

    Spike Timing–Dependent Plasticity: A Hebbian Learning Rule,

    N. Caporale and Y. Dan, “Spike Timing–Dependent Plasticity: A Hebbian Learning Rule,” Annu. Rev. Neurosci., vol. 31, no. 1, pp. 25–46, Jul. 2008, doi: 10.1146/annurev.neuro.31.060407.125639

    doi:10.1146/annurev.neuro.31.060407.125639 2008

  19. [19]

    D. O. Hebb, The organization of behavior; a neuropsychological theory. in The organization of behavior; a neuropsychological theory. Oxford, England: Wiley, 1949, pp. xix, 335

    1949

  20. [20]

    Surrogate Gradient Learning in Spiking Neural Networks: Bringing the Power of Gradient-Based Optimization to Spiking Neural Networks,

    E. O. Neftci, H. Mostafa, and F. Zenke, “Surrogate Gradient Learning in Spiking Neural Networks: Bringing the Power of Gradient-Based Optimization to Spiking Neural Networks,” IEEE Signal Process. Mag., vol. 36, no. 6, pp. 51–63, Nov. 2019, doi: 10.1109/MSP.2019.2931595

    doi:10.1109/msp.2019.2931595 2019

  21. [21]

    The Remarkable Robustness of Surrogate Gradient Learning for Instilling Complex Function in Spiking Neural Networks,

    F. Zenke and T. P. Vogels, “The Remarkable Robustness of Surrogate Gradient Learning for Instilling Complex Function in Spiking Neural Networks,” Neural Computation, vol. 33, no. 4, pp. 899–925, Mar. 2021, doi: 10.1162/neco_a_01367

    doi:10.1162/neco_a_01367 2021

  22. [22]

    Training Deep Spiking Neural Networks Using Backpropagation,

    J. H. Lee, T. Delbruck, and M. Pfeiffer, “Training Deep Spiking Neural Networks Using Backpropagation,” Front. Neurosci., vol. 10, Nov. 2016, doi: 10.3389/fnins.2016.00508

    doi:10.3389/fnins.2016.00508 2016

  23. [23]

    Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing,

    P. U. Diehl, D. Neil, J. Binas, M. Cook, S.-C. Liu, and M. Pfeiffer, “Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing,” in 2015 International Joint Conference on Neural Networks (IJCNN), Killarney, Ireland: IEEE, Jul. 2015, pp. 1–8. doi: 10.1109/IJCNN.2015.7280696

    work page doi:10.1109/ijcnn.2015.7280696 2015

  24. [24]

    Conversion of Continuous-Valued Deep Networks to Efficient Event-Driven Networks for Image Classification,

    B. Rueckauer, I.-A. Lungu, Y. Hu, M. Pfeiffer, and S.- C. Liu, “Conversion of Continuous-Valued Deep Networks to Efficient Event-Driven Networks for Image Classification,” Front. Neurosci., vol. 11, p. 682, Dec. 2017, doi: 10.3389/fnins.2017.00682

    doi:10.3389/fnins.2017.00682 2017

  25. [25]

    Competitive Hebbian learning through spike-timing-dependent synaptic plasticity,

    S. Song, K. D. Miller, and L. F. Abbott, “Competitive Hebbian learning through spike-timing-dependent synaptic plasticity,” Nat Neurosci, vol. 3, no. 9, pp. 919–926, Sep. 2000, doi: 10.1038/78829

    doi:10.1038/78829 2000

  26. [26]

    STDP-based Unsupervised Feature Learning using Convolution- over-time in Spiking Neural Networks for Energy- Efficient Neuromorphic Computing,

    G. Srinivasan, P. Panda, and K. Roy, “STDP-based Unsupervised Feature Learning using Convolution- over-time in Spiking Neural Networks for Energy- Efficient Neuromorphic Computing,” J. Emerg. Technol. Comput. Syst., vol. 14, no. 4, p. 44:1-44:12, Nov. 2018, doi: 10.1145/3266229

    work page doi:10.1145/3266229 2018

  27. [27]

    On-Chip Unsupervised Learning Using STDP in a Spiking Neural Network,

    A. Gupta and S. Saurabh, “On-Chip Unsupervised Learning Using STDP in a Spiking Neural Network,” IEEE Transactions on Nanotechnology, vol. 22, pp. 365–376, 2023, doi: 10.1109/TNANO.2023.3293011

    work page doi:10.1109/tnano.2023.3293011 2023

  28. [28]

    An STDP-Based Supervised Learning Algorithm for Spiking Neural Networks,

    Z. Hu, T. Wang, and X. Hu, “An STDP-Based Supervised Learning Algorithm for Spiking Neural Networks,” vol. 10635, 2017, pp. 92–100. doi: 10.1007/978-3-319-70096-0_10

    work page doi:10.1007/978-3-319-70096-0_10 2017

  29. [29]

    Spatio- Temporal Backpropagation for Training High- Performance Spiking Neural Networks,

    Y. Wu, L. Deng, G. Li, J. Zhu, and L. Shi, “Spatio- Temporal Backpropagation for Training High- Performance Spiking Neural Networks,” Front. Neurosci., vol. 12, p. 331, May 2018, doi: 10.3389/fnins.2018.00331

    doi:10.3389/fnins.2018.00331 2018

  30. [30]

    Surrogate Gradient Learning in Spiking Neural Networks,

    E. O. Neftci, H. Mostafa, and F. Zenke, “Surrogate Gradient Learning in Spiking Neural Networks,” May 03, 2019, arXiv: arXiv:1901.09948. doi: 10.48550/arXiv.1901.09948

    Pith/arXiv arXiv doi:10.48550/arxiv.1901.09948 2019

  31. [31]

    Backpropagation through time: what it does and how to do it,

    P. J. Werbos, “Backpropagation through time: what it does and how to do it,” Proceedings of the IEEE, vol. 78, no. 10, pp. 1550–1560, Oct. 1990, doi: 10.1109/5.58337

    doi:10.1109/5.58337 1990

  32. [32]

    Are we ready for autonomous driving? The KITTI vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? The KITTI vision benchmark suite,” in 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI: IEEE, Jun. 2012, pp. 3354–3361. doi: 10.1109/CVPR.2012.6248074

    doi:10.1109/cvpr.2012.6248074 2012

  33. [33]

    BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning,

    F. Yu et al., “BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning,” Apr. 08, 2020, arXiv: arXiv:1805.04687. doi: 10.48550/arXiv.1805.04687

    Pith/arXiv arXiv doi:10.48550/arxiv.1805.04687 2020

  34. [34]

    Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection,

    X. Li et al., “Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection,” Jun. 08, 2020, arXiv: arXiv:2006.04388. doi: 10.48550/arXiv.2006.04388

    Pith/arXiv arXiv doi:10.48550/arxiv.2006.04388 2020

  35. [35]

    HOTA: A Higher Order Metric for Evaluating Multi-object Tracking,

    J. Luiten et al., “HOTA: A Higher Order Metric for Evaluating Multi-object Tracking,” Int J Comput Vis, vol. 129, no. 2, pp. 548–578, Feb. 2021, doi: 10.1007/s11263-020-01375-2

    doi:10.1007/s11263-020-01375-2 2021

  36. [36]

    BoT- SORT: Robust Associations Multi-Pedestrian Tracking,

    N. Aharon, R. Orfaig, and B.-Z. Bobrovsky, “BoT- SORT: Robust Associations Multi-Pedestrian Tracking,” arXiv.org. Accessed: Nov. 25, 2025. [Online]. Available: https://arxiv.org/abs/2206.14651v2

    Pith/arXiv arXiv 2025

  37. [37]

    C. Han, Q. Zhao, S. Zhang, Y. Chen, Z. Zhang, and J. Yuan, YOLOPv2: Better, Faster, Stronger for Panoptic Driving Perception. 2022. doi: 10.48550/arXiv.2208.11434

    Pith/arXiv arXiv doi:10.48550/arxiv.2208.11434 2022

  38. [38]

    ByteTrack: Multi-Object Tracking by Associating Every Detection Box,

    Y. Zhang et al., “ByteTrack: Multi-Object Tracking by Associating Every Detection Box,” Apr. 07, 2022, arXiv: arXiv:2110.06864. doi: 10.48550/arXiv.2110.06864. Manish Kolachalam was born in Hyderabad, Telangana, India in 1985. He received his B.Sc. degree in biotechnology from Nagpur University, Nagpur, India and his M.Sc. in biotechnology from the Mahara...

    Pith/arXiv arXiv doi:10.48550/arxiv.2110.06864 2022