Dataset Preparation

Notes on Video Data format

MMAction supports two types of data format: raw frames and video. The former is widely used in previous projects such as TSN. This is fast (especially when SSD is available) but fails to scale to the fast-growing datasets. (For example, the newest edition of Kinetics has 650K videos and the total frames will take up several TBs.) The latter save much space but is slower due to video decoding at execution time. To alleviate such issue, we use decord for efficient video loading.

For action recognition, both formats are supported. For temporal action detection and spatial-temporal action detection, we still recommend the format of raw frames.

Supported datasets

The supported datasets are listed below. We provide shell scripts for data preparation under the path $MMACTION/data_tools/. To ease usage, we provide tutorials of data deployment for each dataset.

HMDB51: See PREPARING_HMDB51.md
UCF101: See PREPARING_UCF101.md
Kinetics400: See PREPARING_KINETICS400.md
THUMOS14: See PREPARING_TH14.md
AVA: See PREPARING_AVA.md

Now, you can switch to GETTING_STARTED.md to train and test the model.

TL;DR The following guide is helpful when you want to experiment with custom dataset. Similar to the datasets stated above, it is recommended organizing in $MMACTION/data/$DATASET.

Prepare annotations

Prepare videos

Please refer to the official website and/or the official script to prepare the videos. Note that the videos should be arranged in either (1) a two-level directory organized by ${CLASS_NAME}/${VIDEO_ID} or (2) a single-level directory. It is recommended using (1) for action recognition datasets (such as UCF101 and Kinetics) and using (2) for action detection datasets or those with multiple annotations per video (such as THUMOS14 and AVA).

Extract frames

To extract frames (optical flow, to be specific), dense_flow is needed. (TODO: This will be merged into MMAction in the next version in a smoother way). For the time being, please use the following command:

python build_rawframes.py $SRC_FOLDER $OUT_FOLDER --df_path $PATH_OF_DENSE_FLOW --level {1, 2}

$SRC_FOLDER points to the folder of the original video (for example)
$OUT_FOLDER points to the root folder where the extracted frames and optical flow store
$PATH_OF_DENSE_FLOW points to the root folder where dense_flow is installed.
--level is either 1 for the single-level directory or 2 for the two-level directory

The recommended practice is

set $OUT_FOLDER to be an folder located in SSD
symlink the link $OUT_FOLDER to $MMACTION/data/$DATASET/rawframes.

ln -s ${OUT_FOLDER} $MMACTION/data/$DATASET/rawframes

Generate filelist

cd $MMACTION
python data_tools/build_file_list.py ${DATASET} ${SRC_FOLDER} --level {1, 2} --format {rawframes, videos}

${SRC_FOLDER} should point to the folder of the corresponding to the data format:
- "$MMACTION/data/$DATASET/rawframes" if --format rawframes
- "$MMACTION/data/$DATASET/videos" if --format videos

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!