OMNIVORE is a single vision model for many different visual modalities. It learns to construct representations that are aligned across visual modalities, without requiring training data that specifies correspondences between those modalities. Using OMNIVORE’s shared visual representation, we successfully identify nearest neighbors of left: an image (ImageNet-1K validation set) in vision datasets that contain right: depth maps (ImageNet-1K training set), single-view 3D images (ImageNet-1K training set), and videos (Kinetics-400 validation set).
This repo contains the code to run inference with a pretrained model on an image, video or RGBD image.
Omnivore requires PyTorch and torchvision, please follow PyTorch's getting started instructions for installation. If you are using conda on a linux machine, you can follow the following instructions -
conda create --name omnivore python=3.8
conda activate omnivore
conda install pytorch=1.9.0 torchvision=0.10.0 torchaudio=0.9.0 cudatoolkit=11.1 -c pytorch
We also require einops
, pytorchvideo
and timm
which can be installed via pip -
pip install einops
pip install pytorchvideo
pip install timm
We share checkpoints for all the Omnivore models in the paper. The models are available via torch.hub, and we also share URLs to all the checkpoints.
The details of the models, their torch.hub names / checkpoint links, and their performance is listed below.
Name | IN1k Top 1 | Kinetics400 Top 1 | SUN RGBD Top 1 | Model |
---|---|---|---|---|
Omnivore Swin T | 81.2 | 78.9 | 62.3 | omnivore_swinT |
Omnivore Swin S | 83.4 | 82.2 | 64.6 | omnivore_swinS |
Omnivore Swin B | 84.0 | 83.3 | 65.4 | omnivore_swinB |
Omnivore Swin B (IN21k) | 85.3 | 84.0 | 67.2 | omnivore_swinB_imagenet21k |
Omnivore Swin L (IN21k) | 86.0 | 84.1 | 67.1 | omnivore_swinL_imagenet21k |
Numbers are based on Table 2. and Table 4. in the Omnivore Paper.
The models can be loaded via torch.hub using the following command -
model = torch.hub.load("facebookresearch/omnivore", model="omnivore_swinB")
The class mappings for the datasets can be downloaded as follows:
wget https://s3.amazonaws.com/deep-learning-models/image-models/imagenet_class_index.json
wget https://dl.fbaipublicfiles.com/pyslowfast/dataset/class_names/kinetics_classnames.json
wget https://dl.fbaipublicfiles.com/omnivore/sunrgbd_classnames.json
Follow the inference_tutorial.ipynb
tutorial locally or for step by step instructions on how to run inference with an image, video and RGBD image.
To run the tutorial you need to install jupyter notebook
. For installing it on conda, you may run the follwing:
conda install jupyter nb_conda ipykernel ipywidgets
Integrated into Huggingface Spaces 🤗 using Gradio. Try out the Web Demo
Replicate web demo and docker image is available! You can try loading with different checkpoints here
If this work is helpful in your research, please consider starring ⭐ us and citing:
@article{girdhar2022omnivore,
title={{Omnivore: A Single Model for Many Visual Modalities}},
author={Girdhar, Rohit and Singh, Mannat and Ravi, Nikhila and van der Maaten, Laurens and Joulin, Armand and Misra, Ishan},
journal={arXiv preprint arXiv:2201.08377},
year={2022}
}
We welcome your pull requests! Please see CONTRIBUTING and CODE_OF_CONDUCT for more information.
Omnivore is released under the CC-BY-NC 4.0 license. See LICENSE for additional details. However the Swin Transformer implementation is additionally licensed under the Apache 2.0 license (see NOTICE for additional details).