Pith. sign in

REVIEW 4 major objections 42 references

HOMI: Ultra-Fast EdgeAI platform for Event Cameras

T0 review · 4 major / 0 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read HOMI is an ultra-low-latency edge AI platform that couples a Prophesee IMX636 event camera with a Xilinx Zynq FPGA, and the paper claims it reaches 94% accuracy on DVS Gesture and 1000 fps throughput while using only 33% of available LUT…

desk verdict The HOMI abstract promises a credible event-camera FPGA platform with plausible numbers, but the submitted body is an unrelated VLM paper, so nothing can be verified as-is. read the letter →

arxiv 2508.12637 v1 pith:HLO6TP46 submitted 2025-08-18 cs.AR cs.CVcs.ETcs.NE

Shankaranarayanan H , Satyapreet Singh Yadav , Adithya Krishna , Ajay Vikram P , Mahesh Mehendale , Chetan Singh Thakur This is my paper · ORCID
classification cs.ARcs.CVcs.ETcs.NE
keywords eventcamerasedgeAIFPGAaccelerationDVSGestureultra-lowlatencyevent-basedvisionacceleratorhardware-optimizedpreprocessing
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

HOMI is an end-to-end edge AI platform that combines a Prophesee IMX636 event camera with a Xilinx Zynq UltraScale+ FPGA and a custom AI accelerator. The paper reports that, in its high-accuracy configuration, the platform achieves 94% accuracy on the DVS Gesture dataset, and in its low-latency configuration it sustains 1000 frames per second. The aim is to show that event-based perception can be made both accurate and fast enough for closed-loop control entirely on edge hardware, by exploiting the sparsity and asynchrony of event streams through hardware-optimised preprocessing. A sympathetic reader would care because earlier event-processing systems were typically partial, high-latency, or did not fully exploit event sparsity, and HOMI positions itself as a complete alternative.

What carries the argument

The load-bearing mechanism is the platform itself: a Prophesee IMX636 sensor streaming asynchronous events into a Xilinx Zynq UltraScale+ FPGA that runs an in-house AI accelerator, together with hardware-optimised preprocessing pipelines. These pipelines support two histogram-accumulation modes—constant-time and constant-event—and both linear and exponential time surfaces, converting sparse event streams into regular representations the accelerator can process. The dual-mode design is what lets a single implementation serve both accuracy-driven and low-latency applications, and the claimed 33% LUT usage is the concrete evidence that the design exploits event sparsity rather than over-provisioning compute.

What would settle it

Run HOMI on the full DVS Gesture test set with the commonly used event-window length and all ten classes balanced; if the resulting accuracy falls materially below 94%, or if the 1000 fps throughput cannot be sustained on the stated FPGA without losing events, the central claim would be contradicted.

Watch

Extended reading notes

Core claim

Stated on the paper's own terms, the central claim is that a complete event-camera pipeline—from sensor to pre-processing to a dedicated neural accelerator on the FPGA—can deliver both competitive recognition accuracy and very high throughput on a single edge device. The authors report 94% accuracy on DVS Gesture in a high-accuracy mode and 1000 fps throughput in a low-latency mode, with the entire hardware-optimised pipeline occupying only 33% of the FPGA's look-up tables. That resource figure is part of the claim: it is offered as evidence that the design leaves substantial headroom for more complex models, multi-task deployments, or further latency reduction, rather than consuming the full device.

Load-bearing premise

The headline 94% accuracy assumes the DVS Gesture evaluation used a protocol comparable to prior benchmarks—standard event-window length and balanced classes—but the abstract does not report the protocol.

Editorial extensions

If this is right

  • Gesture-based human-robot interaction and other closed-loop control tasks could run entirely on an event-camera edge device, removing cloud round-trips.
  • The same platform can switch between accuracy-driven and low-latency modes via its dual preprocessing paths, so one hardware design covers different application needs.
  • With only 33% of LUTs used, the remaining FPGA resources can host larger models, additional tasks, or further latency optimisation.
  • The reported figures suggest event-camera edge processing can be competitive with software-based recognizers while running at far higher rates.

Reading between the lines

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

  • Since the preprocessing pipelines are presented as general-purpose, HOMI likely extends to other event-based tasks such as optical flow or object tracking, but the paper only demonstrates DVS Gesture as a use case.
  • If the 94% figure holds under a standard protocol, it would place HOMI at parity with frame-based edge gesture recognizers while gaining the microsecond-level response of event sensors—something frame-based systems cannot match.
  • A natural next measurement the paper leaves implicit is energy per inference, which would determine whether the platform truly fits battery-powered robots.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 0 minor

Summary. The abstract of arXiv:2508.12637 describes HOMI, an end-to-end edge AI platform for event cameras combining a Prophesee IMX636 sensor with a Xilinx Zynq UltraScale+ MPSoC FPGA, and reports 94% accuracy on the DVS Gesture dataset, 1000 fps throughput in a low-latency configuration, and 33% LUT utilization. The manuscript body provided for review, however, is the full text of an unrelated paper, 'edgeVLM: Cloud-edge Collaborative Real-time VLM based on Context Transfer' (arXiv:2508.12638), which addresses vision-language model inference and contains no mention of event cameras, HOMI, the FPGA implementation, or the DVS Gesture experiments. As submitted, the paper provides no methodology, architecture description, evaluation protocol, or comparisons to support the abstract's claims, making them unverifiable from the available text.

Significance. If the abstract's claims are substantiated, HOMI would represent a useful system-level contribution to low-latency edge processing of event-camera data: a commercial sensor-FPGA pairing with hardware optimized preprocessing, support for both constant-time and constant-event histogram accumulation, and linear/exponential time surfaces would have practical value, and the reported 33% LUT utilization suggests headroom for further integration. However, because the submitted body is a different paper entirely, the significance cannot be assessed beyond the abstract. No code, artifacts, or reproducible protocols are provided, and the headline numbers are not connected to any measurable procedure, so the claims are currently unfalsifiable from the submission.

major comments (4)
  1. [Full text (all body sections)] The body of the submission is not the manuscript described by the abstract. All sections, including the references, present the paper 'edgeVLM: Cloud-edge Collaborative Real-time VLM based on Context Transfer' (arXiv:2508.12638), which concerns vision-language model inference and contains no description of the HOMI platform, the Prophesee IMX636 sensor, the Xilinx Zynq FPGA design, the event preprocessing pipelines, or the DVS Gesture experiments. Consequently, the submission provides no technical content by which the abstract's claims for HOMI can be evaluated.
  2. [Abstract] The 94% accuracy on the DVS Gesture dataset is asserted without any evaluation protocol. The abstract does not specify the train/test split, event-window length, histogram mode (constant-time versus constant-event), time-surface decay constants, network architecture, or any comparison with published baselines on this dataset. A bare headline accuracy is not a checkable claim.
  3. [Abstract] The throughput figure of 1000 fps is ambiguous because it is not defined whether this is sustained end-to-end throughput including event sensor readout and preprocessing, or inference-only throughput at a particular event density; the batch size and event density are unspecified. Likewise, the 33% LUT utilization is not anchored to a defined model size, bit width, or configuration, so it cannot be interpreted meaningfully.
  4. [Abstract] The abstract claims the platform is 'end-to-end' and 'ultra-low latency,' but no latency measurements, power measurements, or comparisons to prior event-camera gesture-recognition systems or edge AI accelerators are reported. Without baselines, the claimed accuracy and throughput cannot be placed in context, and the statement that the platform 'cater[s] to both accuracy-driven and low-latency applications' is unsupported by any measurement or ablation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the HOMI claims are external hardware measurements; the body text mismatch raises verifiability concerns but not circular derivation.

full rationale

The HOMI submission consists of an abstract asserting measured outcomes (94% accuracy on DVS Gesture, 1000 fps throughput, 33% LUT utilization) followed by a full text that is actually an unrelated paper, edgeVLM. Within the HOMI portion there are no equations, no fitted parameters, no derivation chain, no self-citations, and no definitional reductions to inspect. The accuracy and throughput figures are presented as empirical results from an external dataset and an FPGA implementation, not as predictions derived from the paper's own inputs. Therefore none of the seven circularity patterns apply. The full-text mismatch is a serious verifiability and presentation defect: the evaluation protocol, model architecture, event-window configuration, train/test split, baseline comparisons, and measurement methodology for the HOMI claims are absent from the supplied text, so the central claims cannot be independently reproduced. However, absence of methodology is a correctness and completeness concern, not a circularity finding. Per the hard rules, circularity may only be asserted when the paper itself exhibits a specific reduction of a claimed result to its inputs; no such reduction is present here. Score is accordingly 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claims rest on unstated preprocessing parameters and the domain assumption that event-camera benchmarks transfer to real robotics. No new physical entities are introduced; the platform is an engineering system.

free parameters (1)
  • Preprocessing parameters (event window, histogram bins, time-surface decay constants)
    The abstract does not report the values used for the DVS Gesture accuracy and 1000 fps throughput; these parameters directly affect both metrics and must be known to reproduce the claims.
assumptions (3)
  • domain assumption Event cameras' asynchronous, sparse output can be exploited by hardware-optimized preprocessing to achieve lower latency than frame-based cameras.
    This is the motivation stated in the abstract; if false, the platform's advantage over conventional vision systems is unclear.
  • domain assumption DVS Gesture dataset accuracy is a meaningful proxy for real-world edge robotics performance.
    The abstract uses DVS Gesture as the demonstration use case; the claim's value depends on this benchmark transferring to deployment.
  • standard math LUT utilization is an appropriate measure of FPGA resource headroom.
    The abstract infers ample headroom from 33% LUT usage; this assumes other resources such as BRAM and DSPs are not the limiting factor, which is not stated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HOMI: Ultra-Fast EdgeAI platform for Event Cameras." pith.science (2026). https://pith.science/paper/HLO6TP46

@misc{pith2026250812637,
  author       = {Pith},
  title        = {Pith review of: HOMI: Ultra-Fast EdgeAI platform for Event Cameras},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HLO6TP46}},
  note         = {Machine review of arXiv:2508.12637}
}
read the original abstract

Event cameras offer significant advantages for edge robotics applications due to their asynchronous operation and sparse, event-driven output, making them well-suited for tasks requiring fast and efficient closed-loop control, such as gesture-based human-robot interaction. Despite this potential, existing event processing solutions remain limited, often lacking complete end-to-end implementations, exhibiting high latency, and insufficiently exploiting event data sparsity. In this paper, we present HOMI, an ultra-low latency, end-to-end edge AI platform comprising a Prophesee IMX636 event sensor chip with an Xilinx Zynq UltraScale+MPSoC FPGA chip, deploying an in-house developed AI accelerator. We have developed hardware-optimized pre-processing pipelines supporting both constant-time and constant-event modes for histogram accumulation, linear and exponential time surfaces. Our general-purpose implementation caters to both accuracy-driven and low-latency applications. HOMI achieves 94% accuracy on the DVS Gesture dataset as a use case when configured for high accuracy operation and provides a throughput of 1000 fps for low-latency configuration. The hardware-optimised pipeline maintains a compact memory footprint and utilises only 33% of the available LUT resources on the FPGA, leaving ample headroom for further latency reduction, model parallelisation, multi-task deployments, or integration of more complex architectures.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 26 canonical work pages

  1. [1]

    Qwen technical report.arXiv preprint arXiv:2309.16609, 2023

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report.arXiv preprint arXiv:2309.16609, 2023. 2

    arXiv 2023

  2. [2]

    Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhao- hai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Jun- yang Lin. Qwen2.5-vl technical repor...

    arXiv 2025

  3. [3]

    Ipn hand: A video dataset and benchmark for real-time continuous hand gesture recog- nition

    Gibran Benitez-Garcia, Jesus Olivares-Mercado, Gabriel Sanchez-Perez, and Keiji Yanai. Ipn hand: A video dataset and benchmark for real-time continuous hand gesture recog- nition. In25th International Conference on Pattern Recog- nition, ICPR 2020, Milan, Italy, Jan 10–15, 2021, pages 1–8. IEEE, 2021. 6

    work page 2020

  4. [4]

    Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom

    Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A mul- timodal dataset for autonomous driving.arXiv preprint arXiv:1903.11027, 2019. 6

    arXiv 1903

  5. [5]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 4

    arXiv 2010

  6. [6]

    Learn- ing a deep convnet for multi-label classification with partial labels

    Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learn- ing a deep convnet for multi-label classification with partial labels. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 647–657, 2019. 6

    work page 2019

  7. [7]

    Notes-guided mllm reasoning: Enhancing mllm with knowl- edge and visual notes for visual question answering

    Wenlong Fang, Qiaofeng Wu, Jing Chen, and Yun Xue. Notes-guided mllm reasoning: Enhancing mllm with knowl- edge and visual notes for visual question answering. InPro- ceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 19597–19607, 2025. 1

    work page 2025

  8. [8]

    History-augmented vision-language models for frontier-based zero-shot object navigation.arXiv preprint arXiv:2506.16623, 2025

    Mobin Habibpour and Fatemeh Afghah. History-augmented vision-language models for frontier-based zero-shot object navigation.arXiv preprint arXiv:2506.16623, 2025. 3

    arXiv 2025

Show all 42 references
  1. [9]

    Laecips: Large vi- sion model assisted adaptive edge-cloud collaboration for iot-based perception system.arXiv e-prints, pages arXiv– 2404, 2024

    Shijing Hu, Ruijun Deng, Xin Du, Zhihui Lu, Qiang Duan, Yi He, Shih-Chia Huang, and Jie Wu. Laecips: Large vi- sion model assisted adaptive edge-cloud collaboration for iot-based perception system.arXiv e-prints, pages arXiv– 2404, 2024. 2, 3, 7

    2024

  2. [10]

    A cloud-edge collaborative architecture for multi- modal llms-based advanced driver assistance systems in iot networks.IEEE Internet of Things Journal, 2024

    Yaqi Hu, Dongdong Ye, Jiawen Kang, Maoqiang Wu, and Rong Yu. A cloud-edge collaborative architecture for multi- modal llms-based advanced driver assistance systems in iot networks.IEEE Internet of Things Journal, 2024. 2, 3, 7

    2024

  3. [11]

    Multi-label classification by exploiting local positive and negative pairwise label correlation.Neurocom- puting, 257:164–174, 2017

    Jun Huang, Guorong Li, Shuhui Wang, Zhe Xue, and Qing- ming Huang. Multi-label classification by exploiting local positive and negative pairwise label correlation.Neurocom- puting, 257:164–174, 2017. 6

    2017

  4. [12]

    Vinci: A real-time embodied smart assistant based on egocentric vision-language model.arXiv preprint arXiv:2412.21080, 2024

    Yifei Huang, Jilan Xu, Baoqi Pei, Yuping He, Guo Chen, Li- jin Yang, Xinyuan Chen, Yaohui Wang, Zheng Nie, Jinyao Liu, et al. Vinci: A real-time embodied smart assistant based on egocentric vision-language model.arXiv preprint arXiv:2412.21080, 2024. 1

    2024 arXiv

  5. [13]

    Gpt-4o system card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perel- man, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli- hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 8

    2024 arXiv

  6. [14]

    In-context learning can perform continual learning like humans.arXiv preprint arXiv:2509.22764, 2025

    Liuwang Kang, Fan Wang, Shaoshan Liu, Hung-Chyun Chou, Chuan Lin, and Ning Ding. In-context learning can perform continual learning like humans.arXiv preprint arXiv:2509.22764, 2025. 4

    2025

  7. [15]

    Generative planning with 3d-vision lan- guage pre-training for end-to-end autonomous driving

    Tengpeng Li, Hanli Wang, Xianfei Li, Wenlong Liao, Tao He, and Pai Peng. Generative planning with 3d-vision lan- guage pre-training for end-to-end autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 4950–4958, 2025. 3

    2025

  8. [16]

    Distributed vlms: Efficient vision-language processing through cloud- edge collaboration

    Yuyang Li, Devika Gumaste, Mehmet Kerem Turkcan, Javad Ghaderi, Gil Zussman, and Zoran Kostic. Distributed vlms: Efficient vision-language processing through cloud- edge collaboration. In2025 IEEE International Conference on Pervasive Computing and Communications Workshops and...

  9. [17]

    Making large vision language mod- els to be good few-shot learners

    Fan Liu, Wenwen Cai, Jian Huo, Chuanyi Zhang, Delong Chen, and Jun Zhou. Making large vision language mod- els to be good few-shot learners. InProceedings of the AAAI Conference on Artificial Intelligence, pages 5415– 5423, 2025. 1

    2025

  10. [18]

    Rea- songrounder: Lvlm-guided hierarchical feature splatting for open-vocabulary 3d visual grounding and reasoning

    Zhenyang Liu, Yikai Wang, Sixiao Zheng, Tongying Pan, Longfei Liang, Yanwei Fu, and Xiangyang Xue. Rea- songrounder: Lvlm-guided hierarchical feature splatting for open-vocabulary 3d visual grounding and reasoning. InPro- ceedings of the Computer Vision and Pattern Recognition...

    2025

  11. [19]

    Bluelm-v-3b: Algorithm and system co-design for multimodal large language models on mobile devices

    Xudong Lu, Yinghao Chen, Cheng Chen, Hui Tan, Boheng Chen, Yina Xie, Rui Hu, Guanxin Tan, Renshou Wu, Yan Hu, et al. Bluelm-v-3b: Algorithm and system co-design for multimodal large language models on mobile devices. In Proceedings of the Computer Vision and Pattern Recognitio...

    2025

  12. [20]

    Can lvlms obtain a driver’s license? a benchmark towards reliable agi for autonomous driving

    Yuhang Lu, Yichen Yao, Jiadong Tu, Jiangnan Shao, Yuexin Ma, and Xinge Zhu. Can lvlms obtain a driver’s license? a benchmark towards reliable agi for autonomous driving. InProceedings of the AAAI Conference on Artificial Intelli- gence, pages 5838–5846, 2025. 1, 3

    2025

  13. [21]

    Maximizing subset accuracy with recur- rent neural networks in multi-label classification.Advances in neural information processing systems, 30, 2017

    Jinseok Nam, Eneldo Loza Menc´ıa, Hyunwoo J Kim, and Jo- hannes F¨urnkranz. Maximizing subset accuracy with recur- rent neural networks in multi-label classification.Advances in neural information processing systems, 30, 2017. 6

    2017

  14. [22]

    Metaspatial: Reinforcing 3d spa- tial reasoning in vlms for the metaverse.arXiv preprint arXiv:2503.18470, 2025

    Zhenyu Pan and Han Liu. Metaspatial: Reinforcing 3d spa- tial reasoning in vlms for the metaverse.arXiv preprint arXiv:2503.18470, 2025. 1

    2025

  15. [23]

    Eve: Efficient multimodal vision lan- 9 guage models with elastic visual experts

    Miao Rang, Zhenni Bi, Chuanjian Liu, Yehui Tang, Kai Han, and Yunhe Wang. Eve: Efficient multimodal vision lan- 9 guage models with elastic visual experts. InProceedings of the AAAI Conference on Artificial Intelligence, pages 6694– 6702, 2025. 1

    2025

  16. [24]

    Lmdrive: Closed-loop end-to-end driving with large language models

    Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120–15130, 2024. 3

    2024

  17. [25]

    Vlm-r1: A stable and generaliz- able r1-style large vision-language model.arXiv preprint arXiv:2504.07615, 2025

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, et al. Vlm-r1: A stable and generaliz- able r1-style large vision-language model.arXiv preprint arXiv:2504.07615, 2025. 8

    2025 arXiv

  18. [26]

    Kitchen scene context based gesture recognition: A contest in icpr2012

    Atsushi Shimada, Kazuaki Kondo, Daisuke Deguchi, G´eraldine Morin, and Helman Stern. Kitchen scene context based gesture recognition: A contest in icpr2012. InInterna- tional Workshop on Depth Image Analysis and Applications, pages 168–185. Springer, 2012. 6

    2012

  19. [27]

    Drivelm: Driving with graph visual question answering

    Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beißwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. InEuropean conference on computer vision, pages 256–274. Springer, 2024. 1

    2024

  20. [28]

    Context-aware llm translation system using conver- sation summarization and dialogue history.arXiv preprint arXiv:2410.16775, 2024

    Mingi Sung, Seungmin Lee, Jiwon Kim, and Sejoon Kim. Context-aware llm translation system using conver- sation summarization and dialogue history.arXiv preprint arXiv:2410.16775, 2024. 3

    2024 arXiv

  21. [29]

    Drivevlm: The convergence of autonomous driving and large vision-language models.arXiv preprint arXiv:2402.12289, 2024

    Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models.arXiv preprint arXiv:2402.12289, 2024. 1, 6

    2024 arXiv

  22. [30]

    Fastvlm: Efficient vision encoding for vision language models

    Pavan Kumar Anasosalu Vasu, Fartash Faghri, Chun-Liang Li, Cem Koc, Nate True, Albert Antony, Gokula Santhanam, James Gabriel, Peter Grasch, Oncel Tuzel, et al. Fastvlm: Efficient vision encoding for vision language models. In Proceedings of the Computer Vision and Pattern Rec...

    2025

  23. [31]

    Hidden in plain sight: Exploring chat history tampering in interactive language models.arXiv preprint arXiv:2405.20234, 2024

    Cheng’an Wei, Yue Zhao, Yujia Gong, Kai Chen, Lu Xi- ang, and Shenchen Zhu. Hidden in plain sight: Exploring chat history tampering in interactive language models.arXiv preprint arXiv:2405.20234, 2024. 3

    2024 arXiv

  24. [32]

    Visual planning: Let’s think only with images.arXiv preprint arXiv:2505.11409, 2025

    Yi Xu, Chengzu Li, Han Zhou, Xingchen Wan, Caiqi Zhang, Anna Korhonen, and Ivan Vuli´c. Visual planning: Let’s think only with images.arXiv preprint arXiv:2505.11409, 2025. 2

    2025

  25. [33]

    Progress-aware video frame captioning

    Zihui Xue, Joungbin An, Xitong Yang, and Kristen Grau- man. Progress-aware video frame captioning. InProceed- ings of the Computer Vision and Pattern Recognition Con- ference, pages 13639–13650, 2025. 1

    2025

  26. [34]

    Minicpm-v: A gpt-4v level mllm on your phone.arXiv preprint arXiv:2408.01800, 2024

    Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone.arXiv preprint arXiv:2408.01800, 2024. 1, 8

    2024 arXiv

  27. [35]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Dar- rell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, ...

    2020

  28. [36]

    Chain-of-focus: Adaptive visual search and zooming for multimodal reasoning via rl.arXiv preprint arXiv:2505.15436, 2025

    Xintong Zhang, Zhi Gao, Bofei Zhang, Pengxiang Li, Xi- aowen Zhang, Yang Liu, Tao Yuan, Yuwei Wu, Yunde Jia, Song-Chun Zhu, et al. Chain-of-focus: Adaptive visual search and zooming for multimodal reasoning via rl.arXiv preprint arXiv:2505.15436, 2025. 2

    2025 arXiv

  29. [37]

    Vavlm: Toward efficient edge-cloud video analytics with vision- language models.IEEE Transactions on Broadcasting, 2025

    Yang Zhang, Hanling Wang, Qing Bai, Haifeng Liang, Pe- ican Zhu, Gabriel-Miro Muntean, and Qing Li. Vavlm: Toward efficient edge-cloud video analytics with vision- language models.IEEE Transactions on Broadcasting, 2025. 3

    2025

  30. [38]

    A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms

    Wangbo Zhao, Yizeng Han, Jiasheng Tang, Zhikai Li, Yib- ing Song, Kai Wang, Zhangyang Wang, and Yang You. A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 19...

    2025

  31. [39]

    Physvlm: Enabling visual language models to understand robotic phys- ical reachability

    Weijie Zhou, Manli Tao, Chaoyang Zhao, Haiyun Guo, Honghui Dong, Ming Tang, and Jinqiao Wang. Physvlm: Enabling visual language models to understand robotic phys- ical reachability. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 6940–6949, 2025. 1

    2025

  32. [40]

    Deepeyes: In- centivizing ”thinking with images” via reinforcement learn- ing

    Zheng Ziwei, Yang Michael, Hong Jack, Zhao Chenxiao, Xu Guohai, Yang Le, Shen Chao, and Yu Xing. Deepeyes: In- centivizing ”thinking with images” via reinforcement learn- ing. 2025. 2

    2025

  33. [41]

    Ground-v: Teaching vlms to ground complex in- structions in pixels

    Yongshuo Zong, Qin Zhang, Dongsheng An, Zhihua Li, Xi- ang Xu, Linghan Xu, Zhuowen Tu, Yifan Xing, and Onkar Dabeer. Ground-v: Teaching vlms to ground complex in- structions in pixels. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 24635–24645,

  34. [286]

    IEEE Computer Society, 2025. 2, 3, 7

    2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.