data

Video data pre-processing

This folder contains the following scripts:

• add_frame_numbering.sh*: draws a huge number on each frame on a specific video;
• dump_data_set.sh*: get images from videos for "traditional training";
• objectify.sh*: convert sampled-data from video-indexed to object-indexed;
• resize_and_split.sh*: see below;
• resize_and_sample.sh*: see below;
• sample_video.sh*: sample the <src video> into k time-subsampled <dst prefix>- videos;
• VideoFolder.py: PyTorch data.Dataset's sub-class for video data loading.

Remove white spaces from file names

White spaces and scripts are not good friends. Replace all spaces with _ in your source data set with the following command, where you have to replace <root data dir> with the correct location.

rename -n "s/ /_/g" <root data dir>/*/*  # for a dry run
rename "s/ /_/g" <root data dir>/*/*     # to rename the files in <root data dir>!!!

In e-VDS35 we have already done this for you. If you plan to use your own data, this step is fundamental!

MatchNet mode

Resize and split videos in train-val

In order to speed up data loading, we shall resize the video shortest side to, say, 256. To do so run the following script.

./resize_and_split.sh <source dir> <destination dir>

By default, the script will

• skip videos shorter than 144 frames (4.8s)
• trim videos longer than 654 frames (21.8s)
• use the last 2s for the validation split
• resize the shortest side to 256px

These options can be varied and turned off by changing the header of the script, which now looks something like this.

################################################################################
# SETTINGS #####################################################################
################################################################################
# comment the next line if you don't want to skip short videos
min_frames=144
# comment the next line if you don't want to limit the max length
max_frames=564
# set split to 0 (seconds) if no splitting is required
split=2
################################################################################

File system

Our input data_set looks somehow like this.

data_set
├── barcode
│   ├── 20160613_140057.mp4
│   ├── 20160613_140115.mp4
│   ├── 20160613_140138.mp4
│   ├── 20160721_023437.mp4
├── bicycle
│   ├── 0707_2_(2).mov
│   ├── 0707_2_(4).mov

Running ./resize_and_sample.sh data_set/ processed-data yields

processed-data/
├── train
│   ├── barcode
│   │   ├── 20160613_140057.mp4
│   │   ├── 20160613_140115.mp4
│   │   ├── 20160613_140138.mp4
│   │   ├── 20160721_023437.mp4
│   ├── bicycle
│   │   ├── 0707_2_(2).mp4
│   │   ├── 0707_2_(4).mp4

TempoNet mode

Resize and sample videos, then splits in train-val

In order to speed up data loading, we shall resize the video shortest side to, say, 256. To do so run the following script.

./resize_and_sample.sh <source dir> <destination dir>

By default, the script will

• skip videos shorter than 144 frames (4.8s)
• trim videos longer than 654 frames (21.8s)
• perform 5 subsamples, use 4 for training, 1 for validation
• resize the shortest side to 256px

These options can be varied and turned off by changing the header of the script, which now looks something like this.

################################################################################
# SETTINGS #####################################################################
################################################################################
# comment the next line if you don't want to skip short videos
min_frames=144
# comment the next line if you don't want to limit the max length
max_frames=564
# set sampling interval: k - 1 train, 1 val
k=5
################################################################################

The output directory will contain as many folders as the total number of videos. Each folder will contain the individual splits.

From video-index- to class-major data organisation

If you would like to train against object classes instead of video indices (like explained in the paper), you also need to run

./objectify.sh <source dir> <destination dir>

This script generates a new directory containing as many folders as classes, filled with symbolic links from the source directory. You can use it for both videos and dumped-videos (images) data sets.

File system

Running ./resize_and_sample.sh data_set/ sampled-data yields

sampled-data/
├── train
│   ├── barcode-20160613_140057
│   │   ├── 1.mp4
│   │   ├── 2.mp4
│   │   ├── 3.mp4
│   │   └── 4.mp4
│   ├── barcode-20160613_140115
│   │   ├── 1.mp4
│   │   ├── 2.mp4

We can "objectify" the structure with ./objectify.sh sampled-data/ object-sampled-data and get

object-sampled-data/
├── train
│   ├── barcode
│   │   ├── 20160613_140057-1.mp4 -> ../../../sampled-data/train/barcode-20160613_140057/1.mp4
│   │   ├── 20160613_140057-2.mp4 -> ../../../sampled-data/train/barcode-20160613_140057/2.mp4
│   │   ├── 20160613_140057-3.mp4 -> ../../../sampled-data/train/barcode-20160613_140057/3.mp4
│   │   ├── 20160613_140057-4.mp4 -> ../../../sampled-data/train/barcode-20160613_140057/4.mp4
│   │   ├── 20160613_140115-1.mp4 -> ../../../sampled-data/train/barcode-20160613_140115/1.mp4
│   │   ├── 20160613_140115-2.mp4 -> ../../../sampled-data/train/barcode-20160613_140115/2.mp4
│   │   ├── 20160613_140115-3.mp4 -> ../../../sampled-data/train/barcode-20160613_140115/3.mp4

To train the discriminative feed-forward branch we need to dump our sampled-data with ./dump_data_set.sh sampled-data/ dumped-sampled-data. The file system will look like this

dumped-sampled-data/
├── train
│   ├── barcode-20160613_140057
│   │   ├── 1001.png
│   │   ├── 1002.png
│   │   ├── 1003.png
│   │   ├── 1004.png
│   │   ├── 1005.png
│   │   ├── 1006.png
│   │   ├── 1007.png

where the "training class" correspond to the video name. If we wish to train against object classes, then we can run ./objectify.sh dumped-sampled-data/ dumped-object-sampled-data and get the following

dumped-object-sampled-data/
├── train
│   ├── barcode
│   │   ├── 20160613_140057-1001.png -> ../../../dumped-sampled-data/train/barcode-20160613_140057/1001.png
│   │   ├── 20160613_140057-1002.png -> ../../../dumped-sampled-data/train/barcode-20160613_140057/1002.png
│   │   ├── 20160613_140057-1003.png -> ../../../dumped-sampled-data/train/barcode-20160613_140057/1003.png
│   │   ├── 20160613_140057-1004.png -> ../../../dumped-sampled-data/train/barcode-20160613_140057/1004.png
│   │   ├── 20160613_140057-1005.png -> ../../../dumped-sampled-data/train/barcode-20160613_140057/1005.png
│   │   ├── 20160613_140057-1006.png -> ../../../dumped-sampled-data/train/barcode-20160613_140057/1006.png
│   │   ├── 20160613_140057-1007.png -> ../../../dumped-sampled-data/train/barcode-20160613_140057/1007.png