pith. sign in

arxiv: 2606.00620 · v1 · pith:CYPNHBW7new · submitted 2026-05-30 · 💻 cs.CV

FlowNar: Scalable Streaming Narration for Long-Form Videos

Zeyun Zhong , Manuel Martin , Chengzhi Wu , David Schneider , Frederik Diederichs , Juergen Gall , Juergen Beyerer This is my paper

Pith reviewed 2026-06-28 19:07 UTC · model grok-4.3

classification 💻 cs.CV
keywords streaming narrationvideo understandingcontext managementlarge multimodal modelsscalable streamingegocentric video
0
0 comments X

The pith

FlowNar uses dynamic visual context removal and a CLAM module to keep memory and computation bounded for streaming narration of arbitrarily long videos.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents FlowNar as a framework designed for narrating long-form videos in real time as they stream. Unlike prior methods whose costs grow with video length, FlowNar removes outdated visual history selectively while retaining key information through its Cross Linear Attentive Memory module. This design ensures both memory usage and processing time stay constant regardless of duration. A new self-conditioned evaluation protocol is introduced to test models under conditions closer to actual deployment. Tests on several egocentric video datasets show improved narration accuracy alongside large gains in efficiency.

Core claim

FlowNar achieves scalable streaming video narration through a dynamic context management strategy that removes historical visual context and a CLAM module for retaining streaming visual history, resulting in bounded visual memory usage and computational complexity.

What carries the argument

dynamic context management strategy for historical visual context removal combined with the CLAM (Cross Linear Attentive Memory) module for streaming visual history retention

If this is right

  • Supports processing videos up to 10 times longer than previous methods without proportional resource increase.
  • Achieves 3 times higher throughput measured in frames per second.
  • Maintains or improves narration quality on Ego4D, EgoExo4D, and EpicKitchens100 datasets compared to baselines.
  • Keeps visual memory usage and computational complexity bounded independent of video duration.

Where Pith is reading between the lines

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

  • Applications such as live sports commentary or assistive devices for the blind could run continuously on limited hardware.
  • The approach might extend to other streaming multimodal tasks like real-time captioning or question answering.
  • Validation against human judgments in live settings would be needed to confirm the evaluation protocol's realism.

Load-bearing premise

Selective removal of historical visual context can be done without losing information essential for correct ongoing narration, and the self-conditioned evaluation protocol reflects real-world deployment conditions.

What would settle it

An experiment showing that for some videos, the narration accuracy drops significantly when key past frames are removed by the dynamic management, or when the self-conditioned protocol scores differ markedly from actual online user feedback.

Figures

Figures reproduced from arXiv: 2606.00620 by Chengzhi Wu, David Schneider, Frederik Diederichs, Juergen Beyerer, Juergen Gall, Manuel Martin, Zeyun Zhong.

Figure 1
Figure 1. Figure 1: Efficiency and scalability comparison of FLOWNAR, Videollm-online (Chen et al., 2024), and Videollm-mod (Wu et al., 2024). (Left) Example streaming outputs: Videollm-online can encounter out of memory (OOM) errors while FLOWNAR continues. (Middle) VRAM usage vs. processed frames (log scale): memory usage for both baselines grows rapidly and exceeds a typical GPU limit (24GB, dashed line), whereas FLOWNAR v… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of FLOWNAR. Streaming video frames vt are encoded and projected, incorporating past memory Mtn−1 at segment start, to produce features Et. Our memory module (CLAM) updates memory tokens Mt using Xt. The LLM processes Et conditioned on cached contexts (C vid t−1, C nar n−1) to trigger the generation of a narration yn. Post-generation, the updated memory Mtn and narration context C nar n are carried… view at source ↗
Figure 3
Figure 3. Figure 3: CLAM mechanism. A gated recurrent state update processes frame tokens {xt,j} sequentially to update state St from St−1. Learnable query vectors Z generate queries Q that read out fixed-size memory tokens Mt from the final state. history from the last completed segment C nar n−1 . If this prob￾ability does not exceed an active threshold θ: p([SKIP] | Et, C vid t−1 , C nar n−1 ) ≤ θ, (1) then the current tim… view at source ↗
Figure 4
Figure 4. Figure 4: Training attention mask. Beyond standard causal mask￾ing, we explicitly block attention (white cells) to raw frames (x) and memories (m) from distant past segments. This enforces re￾liance on the most recent memory tokens and current frames for generating narrations (y), ensuring the model learns to utilize the compressed visual history as required during streaming. beddings {Xt}t∈Sn , the prepended memory… view at source ↗
Figure 5
Figure 5. Figure 5: Examples of FLOWNAR on EpicKitchens100 (Damen et al., 2022). Text in red indicates incorrect narrations [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of the self-conditioned streaming narration process [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Discrepancy regarding positional ids between training and inference [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Wordcloud. We conduct experiments on three challenging, long-form video datasets adapted for the streaming video narration task: Ego4D (Grauman et al., 2022), EgoExo4D (Grauman et al., 2024), and EpicKitchens100 (EK100) (Damen et al., 2022). Key statistics for these datasets, including the number of training/validation samples, average video length, number of segments per video, average segment duration, a… view at source ↗
Figure 11
Figure 11. Figure 11: Speed comparison over sequence length. Our models maintain higher, more stable FPS. C.2. Attention analysis [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Zero-shot generalization to third-person video (ActivityNet). Despite the significant domain shift, the model correctly identifies key semantic activities (e.g., “paint the gate” and “play a game”). Failure case analysis. To understand the limitations of FLOWNAR, we analyze a representative failure case in [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative failure analysis. We observe object hallucinations where the model predicts contextually plausible items rather than the active object, likely due to spatial downsampling (3 × 3) oversmoothing fine details. Additionally, extreme lighting conditions (e.g., 0:34) can result in missed narrations. et al., 2024). Addressing these limitations, the task of streaming video narration (Chen et al., 2024… view at source ↗
read the original abstract

Recent Large Multimodal Models (LMMs), primarily designed for offline settings, are ill-suited for the dynamic requirements of streaming video. While recent online adaptations improve real-time processing, they still face critical scalability challenges, with resource demands typically growing at least linearly with video duration. To overcome this bottleneck, we propose FlowNar, a novel framework for scalable streaming video narration. The core of FlowNar is a dynamic context management strategy for historical visual context removal, combined with our CLAM (Cross Linear Attentive Memory) module for streaming visual history retention, ensuring bounded visual memory usage and computational complexity, crucial for efficient streaming. We also introduce a realistic self-conditioned evaluation protocol and complementary evaluation metrics to assess streaming narration models under deployment-like conditions. Experiments on the Ego4D, EgoExo4D, and EpicKitchens100 datasets demonstrate that FlowNar substantially improves narration quality over strong baselines while being highly efficient, supporting processing of 10$\times$ longer videos and achieving 3$\times$ higher throughput (FPS). The code is available at https://github.com/zeyun-zhong/FlowNar.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes FlowNar, a framework for scalable streaming video narration in LMMs. Its core consists of a dynamic context management strategy for historical visual context removal combined with the CLAM (Cross Linear Attentive Memory) module to enforce bounded visual memory usage and computational complexity. The authors also introduce a self-conditioned evaluation protocol and metrics, and report improved narration quality on Ego4D, EgoExo4D, and EpicKitchens100 while supporting 10× longer videos and 3× higher throughput (FPS). Code is released at the provided GitHub link.

Significance. If the bounded-complexity claim holds under the self-conditioned protocol, the work would address a key scalability barrier for streaming long-form video narration. The release of code and the new evaluation protocol are concrete strengths that facilitate reproducibility and future comparisons.

major comments (2)
  1. [§3.2] §3.2 (Dynamic Context Management): the description of the removal trigger and relevance scoring mechanism is insufficient to verify that long-range dependencies (e.g., delayed references common in Ego4D and EpicKitchens) are preserved; this directly underpins the central bounded-memory claim.
  2. [§4] §4 (Experiments): no ablation isolates the contribution of the dynamic removal policy versus CLAM retention, nor tests failure modes on long-horizon dependencies; without these, the reported gains on the three datasets do not yet substantiate that selective pruning maintains narration accuracy.
minor comments (2)
  1. [§1] The abstract and §1 use “bounded visual memory usage” without an explicit complexity bound (e.g., O(1) or O(log T)); a formal statement would strengthen the claim.
  2. [§3.3] Figure 2 caption and §3.3 notation for CLAM cross-attention could be clarified to distinguish streaming vs. offline modes.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments highlight opportunities to strengthen the exposition of the dynamic context management and to provide more targeted experimental evidence. We address each point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Dynamic Context Management): the description of the removal trigger and relevance scoring mechanism is insufficient to verify that long-range dependencies (e.g., delayed references common in Ego4D and EpicKitchens) are preserved; this directly underpins the central bounded-memory claim.

    Authors: We agree that the current description of the removal trigger and relevance scoring in §3.2 would benefit from greater detail to allow independent verification of long-range dependency preservation. In the revision we will expand this section with explicit mathematical definitions of the trigger condition and scoring function, pseudocode for the removal procedure, and a short discussion with examples drawn from Ego4D and EpicKitchens100 showing how delayed references are retained via the CLAM module. revision: yes

  2. Referee: [§4] §4 (Experiments): no ablation isolates the contribution of the dynamic removal policy versus CLAM retention, nor tests failure modes on long-horizon dependencies; without these, the reported gains on the three datasets do not yet substantiate that selective pruning maintains narration accuracy.

    Authors: While the reported results demonstrate that the full FlowNar system improves quality and efficiency over strong baselines, we acknowledge that the experiments do not yet isolate the dynamic removal policy from CLAM retention or explicitly examine long-horizon failure modes. We will add an ablation study in the revised §4 that compares the complete model against a variant that disables dynamic removal (retaining all history up to the memory bound) and will include a targeted analysis of narration accuracy on video segments containing delayed references. revision: yes

Circularity Check

0 steps flagged

No circularity: new framework and protocol presented as independent contributions

full rationale

The paper's core claims rest on a proposed dynamic context management strategy plus CLAM module for bounded memory in streaming narration, plus a new self-conditioned evaluation protocol. These are introduced as novel elements without any equations or steps that reduce by construction to fitted inputs, self-citations for uniqueness theorems, or renamed known results. The abstract and described contributions contain no self-definitional loops, fitted-input predictions, or load-bearing self-citations; performance claims on Ego4D/EpicKitchens are presented as empirical outcomes rather than tautological. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review based solely on abstract; no explicit free parameters, axioms, or invented entities are described.

pith-pipeline@v0.9.1-grok · 5743 in / 1081 out tokens · 19930 ms · 2026-06-28T19:07:00.585975+00:00 · methodology

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

35 extracted references · 14 canonical work pages · 6 internal anchors

  1. [1]

    Onestory: Co- herent multi-shot video generation with adaptive memory

    An, Z., Jia, M., Qiu, H., Zhou, Z., Huang, X., Liu, Z., Ren, W., Kahatapitiya, K., Liu, D., He, S., et al. Onestory: Co- herent multi-shot video generation with adaptive memory. arXiv preprint arXiv:2512.07802,

    work page arXiv

  2. [2]

    Qwen3-VL Technical Report

    Bai, S., Cai, Y ., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Longformer: The Long-Document Transformer

    Beltagy, I., Peters, M. E., and Cohan, A. Long- former: The long-document transformer.arXiv preprint arXiv:2004.05150,

    work page internal anchor Pith review Pith/arXiv arXiv 2004

  4. [4]

    C., Hampali, S., Sauser, E., Ma, S., et al

    Chatterjee, D., Remelli, E., Song, Y ., Tekin, B., Mittal, A., Bhatnagar, B., Camg¨oz, N. C., Hampali, S., Sauser, E., Ma, S., et al. Memory-efficient streaming videollms for real-time procedural video understanding.arXiv preprint arXiv:2504.13915,

    work page arXiv

  5. [5]

    A simple and effective l 2 norm-based strategy for kv cache compression

    Devoto, A., Zhao, Y ., Scardapane, S., and Minervini, P. A simple and effective l 2 norm-based strategy for kv cache compression. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp. 18476–18499,

    2024

  6. [6]

    Streaming video question- answering with in-context video kv-cache retrieval

    Di, S., Yu, Z., Zhang, G., Li, H., Cheng, H., Li, B., He, W., Shu, F., and Jiang, H. Streaming video question- answering with in-context video kv-cache retrieval. In International Conference on Learning Representations, volume 2025, pp. 42115–42127,

    2025

  7. [7]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces

    Gu, A. and Dao, T. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Lm-infinite: Zero-shot extreme length generalization for large language models

    Han, C., Wang, Q., Peng, H., Xiong, W., Chen, Y ., Ji, H., and Wang, S. Lm-infinite: Zero-shot extreme length generalization for large language models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp. 3991–4008,

    2024

  9. [9]

    Mistral 7B

    URL https: //arxiv.org/abs/2310.06825. Katharopoulos, A., Vyas, A., Pappas, N., and Fleuret, F. Transformers are rnns: Fast autoregressive transformers with linear attention. InInternational conference on ma- chine learning, pp. 5156–5165. PMLR,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Videochat: Chat-centric video understanding.Science China Information Sciences, 68 (10):200102, 2025a

    Li, K., He, Y ., Wang, Y ., Li, Y ., Wang, W., Luo, P., Wang, Y ., Wang, L., and Qiao, Y . Videochat: Chat-centric video understanding.Science China Information Sciences, 68 (10):200102, 2025a. Li, W., Hu, B., Shao, R., Shen, L., and Nie, L. Lion- fs: Fast & slow video-language thinker as online video assistant. InProceedings of the IEEE/CVF Conference on...

    2024

  11. [11]

    Qi, H., Ye, S., Mathis, A., and Mathis, M. W. Llavac- tion: evaluating and training multi-modal large lan- guage models for action understanding.arXiv preprint arXiv:2503.18712,

    work page arXiv

  12. [12]

    Histream: Efficient high- resolution video generation via redundancy-eliminated streaming.arXiv preprint arXiv:2512.21338,

    11 FLOWNAR: Scalable Streaming Narration for Long-Form Videos Qiu, H., Liu, S., Zhou, Z., An, Z., Ren, W., Liu, Z., Schult, J., He, S., Chen, S., Cong, Y ., et al. Histream: Efficient high- resolution video generation via redundancy-eliminated streaming.arXiv preprint arXiv:2512.21338,

    work page arXiv

  13. [13]

    S., Piergiovanni, A., Arnab, A., Dehghani, M., and Angelova, A

    Ryoo, M. S., Piergiovanni, A., Arnab, A., Dehghani, M., and Angelova, A. Tokenlearner: What can 8 learned tokens do for images and videos?arXiv preprint arXiv:2106.11297,

    work page arXiv

  14. [14]

    Keep the cost down: A review on methods to optimize llm’s kv-cache consumption.arXiv preprint arXiv:2407.18003,

    Shi, L., Zhang, H., Yao, Y ., Li, Z., and Zhao, H. Keep the cost down: A review on methods to optimize llm’s kv-cache consumption.arXiv preprint arXiv:2407.18003,

    work page arXiv

  15. [15]

    Retentive Network: A Successor to Transformer for Large Language Models

    Sun, Y ., Dong, L., Huang, S., Ma, S., Xia, Y ., Xue, J., Wang, J., and Wei, F. Retentive network: A successor to transformer for large language models.arXiv preprint arXiv:2307.08621,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Razorattention: Efficient kv cache compression through retrieval heads

    Tang, H., Lin, Y ., Lin, J., Han, Q., Ke, D., Hong, S., Yao, Y ., and Wang, G. Razorattention: Efficient kv cache compression through retrieval heads. InInternational Conference on Learning Representations, volume 2025, pp. 16632–16646,

    2025

  17. [17]

    Effi- cient streaming language models with attention sinks

    Xiao, G., Tian, Y ., Chen, B., Han, S., and Lewis, M. Effi- cient streaming language models with attention sinks. In International Conference on Learning Representations, volume 2024, pp. 21875–21895,

    2024

  18. [18]

    arXiv preprint arXiv:2508.15717 (2025) 5

    Yang, Y ., Zhao, Z., Shukla, S. N., Singh, A., Mishra, S. K., Zhang, L., and Ren, M. Streammem: Query-agnostic kv cache memory for streaming video understanding.arXiv preprint arXiv:2508.15717,

    work page arXiv

  19. [19]

    Video-llama: An instruction- tuned audio-visual language model for video understand- ing

    Zhang, H., Li, X., and Bing, L. Video-llama: An instruction- tuned audio-visual language model for video understand- ing. InProceedings of the 2023 conference on empirical methods in natural language processing: system demon- strations, pp. 543–553, 2023a. Zhang, Z., Sheng, Y ., Zhou, T., Chen, T., Zheng, L., Cai, R., Song, Z., Tian, Y ., R´e, C., Barrett...

    2023

  20. [20]

    Understanding transformer from the perspective of associative memory

    Zhong, S., Xu, M., Ao, T., and Shi, G. Understanding transformer from the perspective of associative memory. arXiv preprint arXiv:2505.19488,

    work page arXiv

  21. [21]

    A Survey on Deep Learning Techniques for Action Anticipation

    Zhong, Z., Martin, M., V oit, M., Gall, J., and Beyerer, J. A survey on deep learning techniques for action anticipation. arXiv preprint arXiv:2309.17257,

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Minigpt- 4: Enhancing vision-language understanding with ad- vanced large language models

    Zhu, D., Shen, X., Li, X., Elhoseiny, M., et al. Minigpt- 4: Enhancing vision-language understanding with ad- vanced large language models. InInternational Con- ference on Learning Representations, volume 2024, pp. 18378–18394,

    2024

  23. [23]

    introduces learnable parameters (MLPs) and benefits from end-to-end training, but it still effectively accumulates information from every incoming frame. When long, 1https://drive.google.com/drive/folders/18i6es_n1RwI4yHJ_6yxvt0MJuEdEa4DO 14 FLOWNAR: Scalable Streaming Narration for Long-Form Videos redundant segments dominate, the memory can become biase...

    2023

  24. [24]

    Dataset statistics Table 8.Dataset statistics

    B.4. Dataset statistics Table 8.Dataset statistics. Dataset # Train # Val Video len. [s] # Segments Seg. dur. # Nar. tokens Ego4D 102.986 17.059 237.7 (±92.0) 53 4.5s 12 EgoExo4D 3.219 826 151.4 (±232.7) 58 2.6s 17 EpicKitchens100 495 138 493.9 (±621.3) 112 4.4s 10 15 FLOWNAR: Scalable Streaming Narration for Long-Form Videos Algorithm 2Self-conditioned s...

    2022

  25. [25]

    Figure 8 further illustrates the most frequent terms found in the narrations for each dataset. B.5. Protocols and metrics Prior streaming narration evaluations (e.g., Videollm-online, Videollm-mod, LION-FS) report quantitative results using ground-truth–conditioned interleaved token sequences constructed from labels (for example [vvnnvv...], where v = fra...

    2019

  26. [26]

    and METEOR (Denkowski & Lavie, 2014)) and structural/semantic metrics such as ROUGE L (Lin,

    2014

  27. [27]

    Each frame is represented by J=10 tokens (1 CLS + 9 spatially averaged 3×3 patch tokens)

    as the visual encoder, processing frames at 2 FPS. Each frame is represented by J=10 tokens (1 CLS + 9 spatially averaged 3×3 patch tokens). A 2-layer MLP projects these visual features from dimension D= 1024 to the LLM’s hidden dimensionDlm = 2048/4096. For the language model, we employ Meta-Llama-3-1B/8B-Instruct (Meta, 2024), adapting all its linear la...

    2048

  28. [28]

    You touch the watch

    for both 1B and 8B language model sizes. The baseline shows a significant drop in FPS as more frames are processed, particularly for the 8B model, due to extensively accumulated context. By employing our dynamic context management strategy to remove historical visual context, FLOWNARmaintains a relatively stable FPS even after 10,000 frames ( 10 FPS for 8...

    2023

  29. [29]

    pick up the knife

    We observe that the model may hallucinate interactions with objects that are visible in the scene or semantically congruent with the environment (e.g., predicting “pick up the knife”) even when they are not being actively manipulated. We attribute this to the aggressive spatial compression (1 CLS +3×3 pooled tokens per frame) required to maintain real-tim...

    2015

  30. [30]

    However, these approaches typically operate in a clip-based manner and explicitly rely on subtitles or dialogue gaps to determine narration timing

    has advanced the field of Movie Audio Description. However, these approaches typically operate in a clip-based manner and explicitly rely on subtitles or dialogue gaps to determine narration timing. Consequently, the majority of these methods operate offline, requiring access to the entire video, and often struggle to scale to long, unsegmented real-world...

    2024

  31. [31]

    We extend Chen et al

    was recently proposed, focusing on generating timely, timestamped descriptions continuously for incoming video segments in an online manner. We extend Chen et al. (2024) by supporting much longer videos and introducing a deployment-like self-conditioned evaluation and metrics. LMMs for online video understanding.Large multimodal models (LMMs) (Alayrac et ...

    2024

  32. [32]

    have significantly advanced multimodal comprehension. Current LMMs address various video understanding benchmarks, including action recognition (Zhao et al., 2023; Qi et al., 2025), temporal action localization (Liu et al., 2024), and video dialogue/question answering (Li et al., 2025a; Song et al., 2024; Maaz et al., 2024; Zhang et al., 2023a; Lin et al....

    2023

  33. [33]

    enforce bounded memory through attention-based KV pruning and event-level tree merging, respectively. Unlike these methods, which produce output only when an external query arrives, FLOWNARmust jointly decidewhenandwhatto narrate over a continuous stream without any prompting, requiring tight coupling between temporal localization and generation under bou...

    2025

  34. [34]

    offers simplicity by retaining only the KV pairs for a fixed window of recent tokens. More sophisticated cache eviction techniques aim to selectively discard less relevant KV pairs based on attention scores (Xiao et al., 2024; Han et al., 2024; Liu et al., 2023b; Zhang et al., 2023b; Adnan et al., 2024), or sparsification (Tang et al., 2025; Yao et al., 2...

    2024

  35. [35]

    (2024) use online K-Means clustering for frame features

    merges similar visual tokens, while Zhou et al. (2024) use online K-Means clustering for frame features. Different from these methods targeting textual caches or using alternative visual compression strategies, we introduce a neural memory mechanism specifically designed to efficiently manage and compress long-term visual context for streaming video analy...

    2024