Pretrained Model on Sportm 1 million about ltc HOT 7 CLOSED

I am planning to release pre-trained models on UCF101, HMDB51 sometime soon. If you need the pre-trained model only on Sports-1M (that corresponds to the 16-frame C3D network by Tran et al.), I could share their model ported to Torch. However, I would need some time to release the modified versions for 100f fine-tuned on UCF101.

from ltc.

@gulvarol
Many Thanks! For now I only need the Torch version of Sport 1M model. It would be a great helpful if you can share it as soon as possible. If I have the Torch version of Sport-1M model, then I can reproduce your experiments. So for now I don't need the 100f and other models. Thanks!

from ltc.

@LongLong-Jing could you try this model? https://lsh.paris.inria.fr/LTC/c3d.t7

from ltc.

@gulvarol Many Thanks! These days I am producing your result with the caffe version of C3D. I have a questions about the Figure 4 in your paper. I tried to reproduce your experiment, and I got the nearly same optical flow clip accuracy. But the optical flow video accuracy of mine is about 3% lower than your figure. This is strange. Can you tell me the exact number of your optical video accuracy in the figure( overall video accuracy)?

from ltc.

@gulvarol We have the same optical clip accuracy but with the different video accuracy. It would be great help if you can tell me both the accuracy of your optical clip and video. Thanks for your help.

from ltc.

@LongLong-Jing yes it's strange. There are many data points in the curve you mention.

These must be overall clip accuracies of flow networks from left to right for 20:20:100 frames:
hf = [80.3100 81.4400 83.2700 82.9000 81.9700]; % high resolution
lf = [77.1600 78.6100 79.0600 80.4700 81.8100]; % low resolution

I am closing this issue as it is about some other question. You can also e-mail me if you want to discuss several different questions.

from ltc.

@gulvarol Thanks for your help. I will test my code again. If I have further questions, I will email you about my questions. Thanks!

from ltc.