chenhao2345

@BishmoyPaul I'm running it on MSRVTT. I have not seen any problems with the fusion score on MSRVTT.

@JosephPai I got similar performance. ~47.5 in stage 1 and 30 in stage 2. It seems to me that the authors get two similarity scores from stage 1 and stage...

According to my experience, the combination of gpu number, batchsize and learning rate can always affect results. I used 4 GPUs to train. Thus, the batchsize and learning rate values...

Hi, Thanks for you question. There is no specific reason. We tried both when we designed this framework. We found that ccloss worked better without "l2 norm". One reason could...

I uploaded examples/test.py. Please modify the following path to load your saved weights. https://github.com/chenhao2345/ICE/blob/30bb04d5457a5b1aaff367461024afd717adc076/examples/test.py#L129

Hi, maybe you can remove Jaccard distance and use Kmeans. Please refer to https://github.com/yxgeee/MMT/blob/057e1ea5d3054c9d7e5fa72c727298d8e4c5f668/examples/mmt_train_kmeans.py

Hi. Yes, you should keep the cross-camera loss. You only need to change the pseudo labels to real labels and keep all the losses.

For example, given a sample from camera 1, Intra-camera loss is the loss between this sample (camera 1) and the same camera center (camera 1); Cross-camera loss is the loss...

@wqydyh I didn't use the intra-camera loss in the training. I keep the code commented out. https://github.com/chenhao2345/ICE/blob/a206eb9a97ad431ab9d9cf38cdcf5ab6fdc6ad1c/ice/trainers.py#L109-L115

Hi. I use a cluster center loss, which already contains intra-camera information. https://github.com/chenhao2345/ICE/blob/a206eb9a97ad431ab9d9cf38cdcf5ab6fdc6ad1c/ice/trainers.py#L151