michigancog/a2cl-pt

Adversarial Background-Aware Loss for Weakly-supervised Temporal Activity Localization (ECCV 2020)

a2cl-pt's Issues

Usage of Beta

Hi,
I did not quite understand how multiplying the original TCAM by a scalar Beta is representing the background feature ? When I debugged it I found out that it is assigning approximately 1/num_clips to each clip as part of the video.

For example if the softmax TCAM for a certain class in 10 number of steps is [0.0839, 0.1689, 0.1689, 0.0767, 0.0798, 0.0798, 0.1025, 0.0767, 0.0798, 0.0831], it's softmax TCAM after multiplying it by beta=0.01 becomes [0.0999, 0.1006, 0.1006, 0.0998, 0.0998, 0.0998, 0.1001, 0.0998, 0.0998, 0.0999].

This is basically assigning equal scores to each clip.

About hard_index

I am sorry to intertrupt you. I have a question about hard index.
In your paper, the hard index follows:

But it uses max function in backward in your code. Why?

forward

A2CL-PT/losses.py

Lines 45 to 47 in 493fb6e

	distHic = distmatH[mask==1]
	distLic = distmatL[mask==1]
	distHicL = torch.min(distmatH[mask==0].view(num_pair, num_class-1), dim=1)[0]

backward

A2CL-PT/losses.py

Lines 66 to 68 in 493fb6e

	distHic = distmatH[mask==1]
	distLic = distmatL[mask==1]
	distHicL, hard_index_batch = torch.max(distmatH[mask==0].view(num_pair, num_class-1), dim=1)

Sharing model weights for Activity Net

Hi,
Would it be possible to share model weights for the activity net training?

ValueError:invalid buffering size when load 'feature_train.npy'

Hi, I meet ValueError:invalid buffering size when load 'feature_train.npy' but it's ok to load 'feature_val.npy'. Is 'feature_train.npy' file broken?

Fail to reproduce the reported performance

Hi,

Thank you very much for your interesting work.

I have run your code with the default parameters, but fail to get the reported results in the paper. Following are the results on validation set:
0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90
My results 0: 26.37 || 98.25 | 56.94, 51.97, 42.12, 33.52, 25.24, 15.52, 7.76, 3.62, 0.66
Reported results 0: 29.99 || 98.39 | 61.17, 56.08, 48.05, 38.98, 30.13, 19.15, 10.55, 4.81, 0.95

Could you please tell me whether I need to modify any other parameters in your codes?

Thank you very much.

Best,
Stacey

some questions about ACLPT_func

Hi,
I have some questions about the implementation of ACLPT_func in losses.py.
(1) line 69, why the index is added 1?

(2) why the gradient need to multiply with xcH?

max snippet length during training

Dear Kyle,
Thanks for your excellent work. It helps me a lot.
I have a question about the maxlen in your code. Do you set maxlen to 200 for both dataset (THUMOS14 and ActivityNet)?
I found the video length in THUMOS14 vary greatly, the length of some videos are even more than 26 minutes (2437 snippets). I am not sure how you set the maxlen for THUMOS14.

Why num_class for THUMOS is 101?

hi @kylemin, when I use you code to train THUMOS14, I print the shape of wtcam, and I found its shape is (batch_size, num_segments, 101), I think it should be 20 instead of 101, since people using the validation set to train network, could you please check this?