search

SilvioGiancola / SoccerNet-code

SoccerNet: A Scalable Dataset for Action Spotting in Soccer Videos
107 stars 25 forks source link

Is there a pre-trained model for quickly test? #10

Closed HymEric closed 4 years ago

juvien commented 4 years ago

this project did you complete?

SilvioGiancola commented 4 years ago

You can find here a copy of the pre-trained model that performed best in the SoccerNet paper.

Note that a novel state-of-the-art has been reached in this CVPR2020 paper, with code to be released soon.

maximelhoustau commented 4 years ago

Hi Silvio,

I tried to use your pre-trained model with my own dataset, specifically on a single video. However when I try to load the models with saver = tf.train.Saver().restore(sess, args.model) in the ClassificationSecondBased.py file, it raises the following error: InvalidArgumentError (see above for traceback): Assign requires shapes of both tensors to match. lhs shape= [64] rhs shape= [512]

I am trying to make a prediction with ResNet features, on a video of 1 minute duration. Features have already been extracted, with a shape of (121, 2048). Would you have any suggestion to make your model work?

SilvioGiancola commented 4 years ago

Hi @maximelhoustau, Have you tried to use the PCA-reduced features for ResNet? Those should be vectors of dimension 512 per image.