search

megvii-research / Iter-E2EDET

Official implementation of the paper "Progressive End-to-End Object Detection in Crowded Scenes"
MIT License
89 stars 8 forks source link

question about lable assignment #3

Closed bravezzzzzz closed 2 years ago

bravezzzzzz commented 2 years ago

Hello, author! Thanks for your work! As described in paper that 'A spatial prior is adopted to compute the matching cost C, that is, the center of bounding box bt needs to fall in the corresponding target box.' I have a few questions.

  1. How is the spatial prior adopted to the matching cost C?
  2. When matching cost absorbs the spatial prior, can each GT match a object query?
  3. What will happen to a certain GT if all the centers of boxes don't fall in it?
    Hoping for your reply!
yexiguafuqihao commented 2 years ago

Hello, author! Thanks for your work! As described in paper that 'A spatial prior is adopted to compute the matching cost C, that is, the center of bounding box bt needs to fall in the corresponding target box.' I have a few questions.

  1. How is the spatial prior adopted to the matching cost C?
  2. When matching cost absorbs the spatial prior, can each GT match a object query?
  3. What will happen to a certain GT if all the centers of boxes don't fall in it? Hoping for your reply!

a.You can check the code in the line of #239 in https://github.com/megvii-research/Iter-E2EDET/blob/main/projects/crowd-e2e-sparse-rcnn/models/matcher.py for the details. And this can still ensure each GT could be matched with only one query.

b. The GT will not participate in the procedure of computing the supervised losses.

liuyang-ict commented 2 years ago

Hi colleagues, I have a little question about the ablation. How does performance change when replacing the original assignment with the hierarchical bipartite matching?

yexiguafuqihao commented 2 years ago

We tried the one-to-one label assignment from a global perspective. Specifically, we first include both reserved queries and noisy ones to perform Hungarian Matching globally to separate positives and negatives. Those reserved queries do not respond to loss computations. Results show comparable performance.

liuyang-ict commented 2 years ago

In my opinion, both global and hierarchical assignments are equivalent. As we all know, Hungarian matching has already considered the class score (corresponding to the class cost in the original source code). Such class cost enables the matcher to identify the reserved and noisy queries.

liuyang-ict commented 2 years ago

Maybe hierarchical assignment pays more attention to the class score, while the global assignment is more focused on localization costs (giou&l1) than class ones.

yexiguafuqihao commented 2 years ago

We design the hierarchical matcher from the intuition that matched targets should be removed from the ground-truths to ensure the noisy queries properly match with the rest ones.