Swathikiran Sudhakaran

1. Due to the smaller size and class imbalance, GTEA61 dataset is a bit tricky. You may try EGTEA dataset which is large enough. 2. The results reported in the...

The feature obtained after spatial average pooling in BNInception is used for the t-SNE.

Yes, you are right.

The output from the backbone (8X2048) is temporal average pooled to obtain a vector (2048) and is used for t-SNE visualization.

bn_inception_gsm.yaml is constructed by adding GSM at the right branches (based on the ablation study reported in the paper) of BNInception. You may compare bn_inception.yaml and bn_inception_gsm.yaml to get a...

Hi, the evaluation code is inside the training code (lines 144-164). You can create an evaluation code based on this, that can load a model and perform inference.

The test set of split2 consists of the videos of user S2. You can create a directory with name test and make a link to the directory S2 inside for...

I see that you are using PyTorchv0.4. I would suggest trying the code with PyTorch0.3.1.

You can either remove the directory or run again with a different directory as output using the option "--outDir ". For GTEA61, four train val splits are generally used. For...

For Diving48, we used a batch size of 8, trained for 20 epochs with a dropout of 0.7 (see Sec. 4.2 in the paper). If you are still getting a...