Test accuracy about c3d HOT 12 CLOSED

No, 44.9 and 46.1 is testing on Sports1M, not UCF101.

But if you training from scratch from UCF101, you will get ~45% (clip accuracy), if you fine-tune the model (which was trained on Sports1M), you will get ~ 77-78% (clip accuracy).

In any case, you can use the model either provided by us (which is pre-trained on Sports1M) or the your own fine-tuned model (on your own data/task) to extract fc6 or prob to make prediction. prob is directly related to the task of the trained net where fc6 can be further train which in our paper using a linear SVM. Hope that helps.

from c3d.

@whjxnyzh this looks OK to me. Note that this is clip-based accuracy. If you want to make video-based accuracy. You can use feature extraction tool, to extract soft-max predictions (e.g. layer "prob" activations) and then aggregate these soft-max predictions to make video-based predictions (and so accuracy). A simple way to aggregate is to average the clip-based predictions to make video-based predictions.

from c3d.

@dutran Thank you very much, i will try it out.

from c3d.

@dutran , Great, "A simple way to aggregate is to average the clip-based predictions to make video-based predictions."
Now,the accuray is 0.8242135871

Thank you very much, this is my first experiment about video using deep learning method

from c3d.

@whjxnyzh @dutran How did you get these results?? When I run train_ucf101.sh it gives me around 45% accuracy!! But you get around 80% accuracy!!
@dutran also in the C3DUserGuide said that we get around 45% accuracy by running train_ucf101.sh!!
I'm confused about the accuracies!! In the paper we have 82% accuracy and here @whjxnyzh get's 82% accuracy but other guys (e.g. #42 ), me and C3DUserGuide reported ~45% accuracy!!
I really appreciate if you clarify this.

Best,

Mohammadreza

from c3d.

@mzolfaghari If you train from scratch, it is supposed to be around 45%. If you fine-tune, you can get higher accuracy 77-78% on clip-accuracy and 82-83% on video-accuracy

from c3d.

Hi again @dutran , I am confused again.... ;( I am reading your article Learning Spatiotemporal Features with 3D Convolutional Networks. In Table 2, C3D(fine-tuned from I380K pre-trained model), the accuracy for one clip hit@1 is 46.1%, not much higher than clip hit@1 of C3D (trained from scratch). Is there something I missed in your article or I misunderstood? Thank you in advance for your help.

from c3d.

There is nothing you missed. As I380K and Sports1M, fine-tuning can help but not making a big difference.

from c3d.

But you said "If you fine-tune, you can get higher accuracy 77-78% on clip-accuracy and 82-83% on video-accuracy". This is a quite big difference...

from c3d.

Yes, you fine-tune, then use fine-tuned model to extract prob, then make predictions. Does that make sense to you?

from c3d.

So this is what I think. Training from scratch on UCF101, then output directly from convnet, the clip-accuracy is around 44.9%. Fine-tuning on UCF101 from the pretrained model on I380K, the clip-accuracy is around 46.1%. But using fine-tuned model, with the extracted features and SVM, we could get 77-78% on clip-accuracy. Is this correct?

from c3d.

Thank you very much

from c3d.