Generic PyTorch Dataset Implementation for Loading, Preprocessing and Augmenting Video Datasets.

If you are completely unfamiliar with loading datasets in PyTorch using and, I recommend
getting familiar with these first through this or

Overview: This example demonstrates the use of VideoFrameDataset

The VideoFrameDataset class serves to easily, efficiently and effectively load video samples from video datasets in PyTorch.

  1. Easily because this dataset class can be used with custom datasets with minimum effort and no modification. The class merely expects the
    video dataset to have a certain structure on disk and expects a .txt annotation file that enumerates each video sample. Details on this
    can be found below and at
  2. Efficiently because the video loading pipeline that this class implements is very fast. This minimizes GPU waiting time during training by eliminating input bottlenecks
    that can slow down training time by several folds.
  3. Effectively because the implemented sampling strategy for video frames is very strong. Video training using the entire sequence of
    video frames (often several hundred) is too memory and compute intense. Therefore, this implementation samples frames evenly from the video (sparse temporal sampling)
    so that the loaded frames represent every part of the video, with support for arbitrary and differing video lengths within the same dataset.
    This approach has shown to be very effective and is taken from
    "Temporal Segment Networks (ECCV2016)" with modifications.

In conjunction with PyTorch's DataLoader, the VideoFrameDataset class returns video batch tensors of size BATCH x FRAMES x CHANNELS x HEIGHT x WIDTH.

For a demo, visit

QuickDemo (

root = os.path.join(os.getcwd(), 'demo_dataset')  # Folder in which all videos lie in a specific structure
annotation_file = os.path.join(root, 'annotations.txt')  # A row for each video sample as: (VIDEO_PATH NUM_FRAMES CLASS_INDEX)

dataset = VideoFrameDataset(

sample = dataset[0]  # take first sample of dataset 
frames = sample[0]   # list of PIL images
label = sample[1]    # integer label

for image in frames:


1. Requirements

# Without these three, VideoFrameDataset will not work.
torchvision >= 0.8.0
torch >= 1.7.0
python >= 3.6

2. Custom Dataset

To use any dataset, two conditions must be met.

  1. The video data must be supplied as RGB frames, each frame saved as an image file. Each video must have its own folder, in which the frames of
    that video lie. The frames of a video inside its folder must be named uniformly as img_00001.jpg ... img_00120.jpg, if there are 120 frames. The filename template
    for frames is then "img_{:05d}.jpg" (python string formatting, specifying 5 digits after the underscore), and must be supplied to the
    constructor of VideoFrameDataset as a parameter. Each video folder lies inside a root folder of this dataset.
  2. To enumerate all video samples in the dataset and their required metadata, a .txt annotation file must be manually created that contains a row for each
    video sample in the dataset. The training, validation, and testing datasets must have separate annotation files. Each row must be a space-separated list that contains
    VIDEO_PATH NUM_FRAMES CLASS_INDEX. The VIDEO_PATH of a video sample should be provided without the root prefix of this dataset.

This example project demonstrates this using a dummy dataset inside of demo_dataset/, which is the root dataset folder of this example. The folder
structure looks as follows:

├───jumping # arbitrary class folder naming
│       ├───0001  # arbitrary video folder naming
│       │     ├───img_00001.jpg
│       │     .
│       │     └───img_00017.jpg
│       └───0002
│             ├───img_00001.jpg
│             .
│             └───img_00018.jpg
└───running # arbitrary folder naming
        ├───0001  # arbitrary video folder naming
        │     ├───img_00001.jpg
        │     .
        │     └───img_00015.jpg


The accompanying annotation .txt file contains the following rows

jumping/0001 17 0
jumping/0002 18 0
running/0001 15 1
running/0002 15 1

Instantiating a VideoFrameDataset with the root_path parameter pointing to demo_dataset, the annotationsfile_path parameter pointing to the annotation file, and
the imagefile_template parameter as "img_{:05d}.jpg", is all that it takes to start using the VideoFrameDataset class.

3. Video Frame Sampling Method

When loading a video, only a number of its frames are loaded. They are chosen in the following way:

  1. The frame indices [1,N] are divided into NUM_SEGMENTS even segments. From each segment, FRAMES_PER_SEGMENT consecutive indices are chosen at random.
    This results in NUM_SEGMENTS*FRAMES_PER_SEGMENT chosen indices, whose frames are loaded as PIL images and put into a list and returned when calling


4. Using VideoFrameDataset for training

As demonstrated in, we can use PyTorch's class with VideoFrameDataset to take care of shuffling, batching, and more.
To turn the lists of PIL images returned by VideoFrameDataset into tensors, the transform video_dataset.imglist_totensor() can be supplied
as the transform parameter to VideoFrameDataset. This turns a list of N PIL images into a batch of images/frames of shape N x CHANNELS x HEIGHT x WIDTH.
We can further chain preprocessing and augmentation functions that act on batches of images onto the end of imglist_totensor().

As of torchvision 0.8.0, all torchvision transforms can now also operate on batches of images, and they apply deterministic or random transformations
on the batch identically on all images of the batch. Therefore, any torchvision transform can be used here to apply video-uniform preprocessing and augmentation.

Pytorch transforms are applied to individual dataset samples (in this case a video frame PIL list, or a frame tensor after imglist_totensor()) before
batching. So, any transforms used here must expect its input to be a frame tensor of shape FRAMES x CHANNELS x HEIGHT x WIDTH or a list of PIL images if imglist_totensor() is not used.

5. Conclusion

A proper code-based explanation on how to use VideoFrameDataset for training is provided in

6. Acknowledgements

We thank the authors of TSN for their codebase, from which we took VideoFrameDataset and adapted it
for general use and compatibility.

    title={Temporal Segment Networks: Towards Good Practices for Deep Action Recognition},
    author={Limin Wang and Yuanjun Xiong and Zhe Wang and Yu Qiao and Dahua Lin and
            Xiaoou Tang and Luc {Val Gool}},
    booktitle={The European Conference on Computer Vision (ECCV)},