search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.39k stars 652 forks source link

Some question about color thr in BoxInst #547

Open Yangr116 opened 2 years ago

Yangr116 commented 2 years ago

Thanks for your great work! In the paper, Table1(a) shows that the color similarity threshold of 0.1 catches the best performance. However, it is set to 0.3 in the configs(line 25). I would like to know what color similarity threshold could obtain the best performance.

Yangr116 commented 2 years ago

Another question about BoxInst. In the pairwise term, the the probability of making the same prediction = p_i p_j + (1 - p_i) (1 - p_j). However, lines 41-45 in 'adet/modeling/condinst/dynamic_mask_head.py' use the max to calculate this probability. 

max_ = torch.max(log_same_fg_prob, log_same_bg_prob)
log_same_prob = torch.log(torch.exp(log_same_fg_prob - max_) +torch.exp(log_same_bg_prob - max_)) + max_

Why does not use

log_same_prob = torch.log(torch.exp(log_same_fg_prob)+torch.exp(log_same_bg_prob))

, directly? Looking forward to your reply~

fxooooooooo commented 2 years ago

Have you solved the above problems? I have the same questions.

engrjav commented 2 years ago

@xulinxulin same question

Yangr116 commented 2 years ago

@engrjav @engrjav From my experiments, the default settings are best. However, the second question is still not solved.

engrjav commented 2 years ago

@Yangr116 i am facing an issue in training, which i posted as Issue no 572 (https://github.com/aim-uofa/AdelaiDet/issues/572) i get zero AP after training my custom dataset for 90 k iterations with the error No weights in checkpoint matched with model. Some model parameters or buffers are not found in the checkpoint:

can you guide