Welcome to join the the Large-Scale Video Classification Challenge workshop in conjunction with ACM Multimedia 2017!
This toolkit contains metadata of this dataset and a python script to compute the mean Average Precision across all 500 categories. Details are listed below.
- lsvc_train.txt, contains the ids of training videos and their corresponding labels.
- lsvc_val.txt, contains the ids of validation videos and their corresponding labels.
- lsvc_test.txt, contains the ids of testing videos.
- lsvc_class_index.txt, contains the class indices for 500 categories.
We provide pre-extracted CNN features based on the fc6 layer of a VGG19 model pretrained on the full ImageNet data. A frame-level feature file (lsvcxxxxxx_fc6_vgg19_frame.binary) and a video-level feature file (lsvcxxxxxx_fc6_vgg19_video.binary) are associated with each video clip. The video-level features are generated by average pooling accross frames.
The feature files are stored in binary format. Here is an example of how to read the features with numpy: frame_features = numpy.fromfile('lsvcxxxxxx_fc6_vgg19_frame.binary', dtype='float32').reshape(-1, 4096)
You will then get a 2D matrix whose size is N*4096, where N is the number of frames for this video and the i-th row of this matrix corresponds to the i-th frame of the video.
Example usage: python compute_average_precision.py --labels samples/lsvc_val_sample.txt --predicted samples/val_result_sample.txt