8000 Question Regarding the "Dynamic Template Capture Module" Implementation · Issue #5 · GXNU-ZhongLab/DUTrack · GitHub
[go: up one dir, main page]

Skip to content

Question Regarding the "Dynamic Template Capture Module" Implementation #5

Open

Description

@vero1925

Dear authors,

First of all, thank you for sharing this interesting work — the design of the Dynamic Template and Language module is very insightful and creative.

While reading through your codebase, I noticed that the dynamic template features — specifically aux_dict['attn_l2s'] from the backbone — do not seem to be used during inference. Instead, the template update mechanism used during inference appears to be different, and I could not find a description of it in the paper.

I would like to ask if this discrepancy might be due to a code version issue. If so, would it be possible to update the repository with the correct implementation? Alternatively, if I have misunderstood the intended design, please feel free to correct me.

Thank you in advance for your clarification and your excellent work!

Activity

  1. cyz2022 commented on Apr 22, 2025

    @cyz2022

    文本动态更新模块似乎和论文中提及的也不匹配,在dutrack.py 中的 ifupdata直接是return true并没有按照论文中所说的比较更新来处理,每一帧都会刷新文本。

  2. xiaohaiLi12 commented on Apr 23, 2025

    @xiaohaiLi12
    Member

    文本动态更新模块似乎和论文中提及的也不匹配,在dutrack.py 中的 ifupdata直接是return true并没有按照论文中所说的比较更新来处理,每一帧都会刷新文本。

    论文要仔细看,消融实验有说逐帧更新的效果是最好的,另外if updata 用来控制更新频率,可以自己去探索什么样的数值更合适。

  3. xiaohaiLi12 commented on Apr 23, 2025

    @xiaohaiLi12
    Member

    Dear authors,

    First of all, thank you for sharing this interesting work — the design of the Dynamic Template and Language module is very insightful and creative.

    While reading through your codebase, I noticed that the dynamic template features — specifically aux_dict['attn_l2s'] from the backbone — do not seem to be used during inference. Instead, the template update mechanism used during inference appears to be different, and I could not find a description of it in the paper.

    I would like to ask if this discrepancy might be due to a code version issue. If so, would it be possible to update the repository with the correct implementation? Alternatively, if I have misunderstood the intended design, please feel free to correct me.

    Thank you in advance for your clarification and your excellent work!

    l2s is stored in the temporal_query field of the aux dictionary, and is used in dutrack.py to update the value of self.track_query for inference in the next frame.”

  4. vero1925 commented on Apr 23, 2025

    @vero1925
    Author

    Thank you very much for your response!

    To follow up, could you kindly point out the exact file path and line number where aux_dict['attn_l2s'] is used during inference? I tried searching for "attn_l2s" across the entire project but could not locate any relevant usage, including in dutrack.py.

    In addition, I would also appreciate it if you could confirm the current template update strategy implemented in lib/test/tracker/dutrack.py. Based on my understanding of the code, the current approach seems to update the template by cropping the newly predicted target region frame-by-frame, storing it as a historical template, and then using these directly as input in subsequent inference steps. This appears to differ from the method described in the paper, where a dynamic template mechanism is involved.

    Please let me know if I’ve misunderstood any part of the design. Looking forward to your clarification, and thank you again for your kind help!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions

      0