Video-FocalNets

Video-FocalNet is an architecture for efficient video recognition that is effectively modeled on both local and global contexts. A spatio-temporal focal modulation approach is utilized, in which self-attention steps are optimized for greater efficiency through cost-effective convolution and element-wise multiplication. After extensive exploration, the parallel spatial and temporal encoding was determined to be the best design choice.

This is a unofficial Keras implementation of Video-FocalNet. The official PyTorch implementation is here

News

• [02-11-2023]: Add gradio web app, huggingface space.
• [31-10-2023]: GPU(s), TPU-VM for fine-tune training are supported, colab.
• [31-10-2023]: Video-FocalNet checkpoints for Driving-48 becomes available, link.
• [30-10-2023]: Video-FocalNet checkpoints for ActivityNet becomes available, link.
• [29-10-2023]: Video-FocalNet checkpoints for SSV2 becomes available, link.
• [28-10-2023]: Video-FocalNet checkpoints for Kinetics-600 becomes available, link.
• [27-10-2023]: Video-FocalNet checkpoints for Kinetics-400 becomes available, link.
• [27-10-2023]: Code of Video-FocalNet in Keras becomes available.

Install

git clone https://github.com/innat/Video-FocalNets.git
cd Video-FocalNets
pip install -e . 

Usage

The Video-FocalNet checkpoints are available in both SavedModel and H5 formats. The variants of this models are tiny, small, and base. Check this release and model zoo page to know details of it. Following are some hightlights.

Inference

>>> from videofocalnet import VideoFocalNetT
>>> model = VideoFocalNetT(name='FocalNetT_K400')
>>> _ = model(np.ones(shape=(1, 8, 224, 224, 3)))
>>> model.load_weights('TFVideoFocalNetT_K400_8x224.h5')
>>> container = read_video('sample.mp4')
>>> frames = frame_sampling(container, num_frames=8)
>>> y = model(frames)
>>> y.shape
TensorShape([1, 400])

>>> probabilities = tf.nn.softmax(y_pred_tf)
>>> probabilities = probabilities.numpy().squeeze(0)
>>> confidences = {
    label_map_inv[i]: float(probabilities[i]) \
    for i in np.argsort(probabilities)[::-1]
}
>>> confidences

A classification results on a sample from Kinetics-400.

Video	Top-5
	{ 'playing cello': 0.8959084749221802, 'playing violin': 0.023411624133586884, 'playing recorder': 0.0011349919950589538, 'playing piano': 0.00101949623785913, 'playing clarinet': 0.0009982039919123054 }

Fine Tune

Each video-focalnet checkpoints returns logits. We can just add a custom classifier on top of it. For example:

# import pretrained model, i.e.
video_focalnet = keras.models.load_model(
    'TFVideoFocalNetB_K400_8x224', compile=False
    )
video_focalnet.trainable = False

# downstream model
model = keras.Sequential([
    video_focalnet,
    layers.Dense(
        len(class_folders), dtype='float32', activation=None
    )
])
model.compile(...)
model.fit(...)
model.predict(...)

Spatio-Temporal Modulator [GradCAM]

Here are some visual demonstration of first and last layer Spatio-Temporal Modulator of Video-FocalNet. More details visual-gradcam.ipynb.

video_sample1.mp4

Model Zoo

The 3D video-focalnet checkpoints are listed in MODEL_ZOO.md.

TODO

• Custom fine-tuning code.
• Support Keras V3 to support multi-framework backend.
• Publish on TF-Hub.

Citation

If you use this video-focalnet implementation in your research, please cite it using the metadata from our CITATION.cff file.

@InProceedings{Wasim_2023_ICCV,
    author    = {Wasim, Syed Talal and Khattak, Muhammad Uzair and Naseer, Muzammal and Khan, Salman and Shah, Mubarak and Khan, Fahad Shahbaz},
    title     = {Video-FocalNets: Spatio-Temporal Focal Modulation for Video Action Recognition},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year      = {2023},
}