The AVA Dataset is a newly exciting dataset for action detection and localization.

I will put a few tools here that may be useful for future development.

1. FFMPEG
2. OpenCV

Following CVDF's latest instruction, we are able to download the AVA v2.1 train/val videos with their urls.

$ cd ava-dataset-tool/video/trainval
$ ./download.sh

Download the test videos with their urls.

$ cd ava-dataset-tool/video/test
$ ./download.sh

The trainval annotations can be downloaded by

$ cd ava-dataset-tool
$ wget https://s3.amazonaws.com/ava-dataset/annotations/ava_v2.1.zip
$ unzip ava_v2.1.zip

Data preprocessing is important for network training. Previously I tried OpenCV, but it is challenging to extract the exact keyframe maybe due to different codec settings. Currently we extract the video clips and the keyframes using ffmpeg. We visualize the bboxes for each keyframe in a new folder preproc. Note that this script extracts keyframes, bboxes, and the 3-second video clips from training set.

$ cd ava-dataset-tool
$ python3 extract_keyframe.py

If you would like to extract video clips and keyframes from the validation set or test set, you may need slightly adjust the corresponding path in extract_keyframe.py.

Stay tuned..