Hi,

I'm trying to reproduce the results reported in the paper using this repo. But the value for action classification on Thumos14 dataset is somewhat not matching the reported result (I am getting 71.67 mAP instead of 86.9 reported in the paper). I have use all the default parameters. Is there something I am missing?

Other values are somewhat close to the values reported.

I am waiting for your reply. Thanks in advance.

so where is the .npy ?

the
if (labels[i] > 0).sum() == 0 or ((labels[i] > 0).sum() != 1 and itr < itr_th): continue
in

def CENTERLOSS(features, logits, labels, seq_len, criterion, itr, device):
    lab = torch.zeros(0).to(device)
    feat = torch.zeros(0).to(device)
    itr_th = 5000    
    for i in range(features.size(0)):
        if (labels[i] > 0).sum() == 0 or ((labels[i] > 0).sum() != 1 and itr < itr_th):
            continue
        # categories present in the video
        labi = torch.arange(labels.size(1))[labels[i]>0]
        atn = F.softmax(logits[i][:seq_len[i]], dim=0)
        atni = atn[:,labi]
        # aggregate features category-wise
        for l in range(len(labi)):
            labl = labi[[l]].float()
            atnl = atni[:,[l]]
            atnl[atnl<atnl.mean()] = 0
            sum_atn = atnl.sum()
            if sum_atn > 0:
                atnl = atnl.expand(seq_len[i],features.size(2))
                # attention-weighted feature aggregation
                featl = torch.sum(features[i][:seq_len[i]]*atnl,dim=0,keepdim=True)/sum_atn
                feat = torch.cat([feat, featl], dim=0)
                lab = torch.cat([lab, labl], dim=0)
        
    if feat.numel() > 0:
        # Compute loss
        loss = criterion(feat, lab)
        return loss / feat.size(0)
    else:
        return 0

Does it mean center-loss use on multi-label training video?

Hi,

Is it possible to provide us with the code for reproducing the figure 4 from the paper?

Nice work and congrats to your ICCV paper. Thanks for sharing your code.

As you mentioned in the paper, for THUMOS14, you follow the setting in STPN(CVPR 18): use 200 videos(20 categories) in the val set for training.

I did not carefully go through every line of your code. It seems that you use all 1010 videos in the val set to train your classifier, this is fair for the action classification task. But it seems that you use the same network to perform the temporal action localization task. I don't think this is the standard protocol for weakly-supervised temporal action localization.

How to make prediction of a video? What is the threshold you choose usually? I am talking about the following line in the paper

After training the 3C-Net, the CLS module (see Fig. 2
and Eq. 2) is used to compute the action-class scores (pmf)
at the video-level using the final T-CAM, for the action classification task

Hi, Sanath

I ran the code recently using I3D feature and got the same result reported in the paper with [email protected] = 26.6 on THUMOS14 dataset.
However, I saved the predictions and used the official code for evaluation, I got 23.
Are the results reported in the paper using the evaluation code in detectionMAP.py or the official code?
Could please release the model checkpoint if you are using the official code?

Thank you
Fan

I run the code for a several times, the mAP at IOU=0.5 is 25.34%. Then I use exponential lr scheduling, and it increases to 26.06%. But still lower than 26.6% as reported. Can you possibly explain what could be wrong? Thanks in advance.