🔥【AAAI'2023, IJCV'2023】Revisiting Classifier: Transferring Vision-Language Models for Video Recognition

Wenhao Wu^1,2, Zhun Sun², Wanli Ouyang^3,1

¹The University of Sydney, ²Baidu, ³Shanghai AI Lab

---

This is the official implementation of the AAAI paper Revisiting Classifier: Transferring Vision-Language Models for Video Recognition, and IJCV paper Transferring Vision-Language Models for Visual Recognition: A Classifier Perspective.

🙋 I also have other cross-modal video projects that may interest you ✨.

Bidirectional Cross-Modal Knowledge Exploration for Video Recognition with Pre-trained Vision-Language Models
Wenhao Wu, Xiaohan Wang, Haipeng Luo, Jingdong Wang, Yi Yang, Wanli Ouyang

Cap4Video: What Can Auxiliary Captions Do for Text-Video Retrieval?
Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, Wanli Ouyang
Accepted by CVPR 2023 as 🌟Highlight🌟 |

📣 Updates

• Aug 07, 2023 The extension of Text4Vis has been accepted by International Journal of Computer Vision (IJCV).
• Dec 22, 2022 Models: The pre-trained models & logs.
• Nov 30, 2022 Config: All the configs (general/few-shot/zero-shot video recognition) on Kinetics-400 & 600, ActivityNet, UCF, and HMDB.
• Nov 30, 2022 Code: Zero-shot Evaluation: Half-classes evaluation and Full-classes evaluation.
• Nov 28, 2022 Code: Single-Machine/Multi-Machine Multi-GPU Distributed Training, Distributed testing.
• Nov 19, 2022 🎉Our paper has been accepted by AAAI-2023.
• Jul 1, 2022 💡Our initial Arxiv paper is released.

🌈 Overview

In our Text4Vis, we revise the role of the linear classifier and replace the classifier with the different knowledge from pre-trained model. We utilize the well-pretrained language model to generate good semantic target for efficient transferring learning.

Content

• Prerequisites
• Data Preparation
• Model Zoo
• Training
• Testing
• BibTeX & Citation
• Acknowledgment

📕 Prerequisites

The code is built with following libraries:

• PyTorch >= 1.8
• RandAugment
• pprint
• tqdm
• dotmap
• yaml
• csv
• Optional: decord (for on-the-fly video training)
• Optional: torchnet (for mAP evaluation on ActivityNet)

📚 Data Preparation

Video Loader

(Recommend) To train all of our models, we extract videos into frames for fast reading. Please refer to MVFNet repo for the detaied guide of data processing.
The annotation file is a text file with multiple lines, and each line indicates the directory to frames of a video, total frames of the video and the label of a video, which are split with a whitespace. Here is the format:

abseiling/-7kbO0v4hag_000107_000117 300 0
abseiling/-bwYZwnwb8E_000013_000023 300 0

(Optional) We can also decode the videos in an online fashion using decord. This manner should work but are not tested. All of the models offered have been trained using offline frames. Example of annotation:

abseiling/-7kbO0v4hag_000107_000117.mp4 0
abseiling/-bwYZwnwb8E_000013_000023.mp4 0

Annotation

Annotation information consists of two parts: video label, and category description.

• Video Label: As mentioned above, this part is same as the traditional video recognition. Please refer to lists/k400/kinetics_rgb_train_se320.txt for the format.
• Category Description: We also need a textual description for each video category. Please refer to lists/kinetics_400_labels.csv for the format.

📱 Model Zoo

Here we provide some off-the-shelf pre-trained checkpoints of our models in the following tables.

#Frame = #input_frame x #spatial crops x #temporal clips

Kinetics-400

Architecture	#Frame	Top-1 Acc.(%)	checkpoint	Train log	config
ViT-B/32	8x3x4	80.0	Github	log	config
ViT-B/32	16x3x4	80.5	Github	log	config
ViT-B/16	8x3x4	82.9	Github	log	config
ViT-B/16	16x3x4	83.6	Github	log	config
ViT-L/14*	8x3x4	86.4	OneDrive	log	config
ViT-L/14-336	8x3x4	87.1	OneDrive	log	config
ViT-L/14-336	32x3x1	87.8	OneDrive	log	config

Note: * indicates that this ViT-L model is used for the zero-shot evaluation on UCF, HMDB, ActivityNet and Kinetics-600.

ActivityNet

Architecture	#Frame	mAP (%)	checkpoint	Train log	config
ViT-L/14	16x1x1	96.5	OneDrive	config
ViT-L/14-336	16x1x1	96.9	OneDrive	log	config

UCF-101

Architecture	#Frame	Top-1 Acc. (%)	checkpoint	Train log	config
ViT-L/14	16x1x1	98.1	OneDrive	log	config

HMDB-51

Architecture	#Frame	Top-1 Acc. (%)	checkpoint	Train log	config
ViT-L/14	16x1x1	81.3	OneDrive	log	config

🚀 Training

This implementation supports Multi-GPU DistributedDataParallel training, which is faster and simpler than DataParallel used in ActionCLIP.

• Single Machine: To train our model on Kinetics-400 with 8 GPUs in Single Machine, you can run:

# For example, train the 8 Frames ViT-B/32.
sh scripts/run_train.sh  configs/k400/k400_train_rgb_vitb-32-f8.yaml

• Mulitple Machines: We also provide the script to train larger model with Mulitple Machines (e.g., 2 machines and 16 GPUs), you can run:

# For example, we train the 8 Frames ViT-L/14 with 2 machines as follows:
# For first machine, you need to set the ip of your first machine as the --master_addr, --nnodes is 2.
# Compared with the Single-Machine training script, only one node_id needs to be added.
sh scripts/run_train_multinodes.sh configs/k400/k400_train_rgb_vitl-14-f8.yaml 0

# For second machine, --master_addr is still the ip of your first machine
sh scripts/run_train_multinodes.sh configs/k400/k400_train_rgb_vitl-14-f8.yaml 1

• Few-shot Recognition: To train our model under Few-shot scenario, you just need to add one line in the general config file:

# You can refer to config/k400/k400_few_shot.yaml
data: 
    ...  # general configurations
    shot: 2  # i.e., 2-shot setting

⚡ Testing

We support single view validation and multi-view (4x3 views) validation.

General/Few-shot Video Recognition

# Single view evaluation. e.g., ViT-B/32 8 Frames on Kinetics-400
sh scripts/run_test.sh  configs/k400/k400_train_rgb_vitb-32-f8.yaml exp/k400/ViT-B/32/f8/last_model.pt

# Multi-view evalition (4clipsx3crops). e.g., ViT-B/32 8 Frames on Kinetics-400
sh scripts/run_test.sh  configs/k400/k400_train_rgb_vitb-32-f8.yaml exp/k400/ViT-B/32/f8/last_model.pt --test_crops 3  --test_clips 4

Zero-shot Evaluation

We use the Kinetics-400 pre-trained model (e.g., ViT-L/14 with 8 frames) to perform cross-dataset zero-shot evaluation, i.e., UCF101, HMDB51, ActivityNet, Kinetics-600.

• Half-classes Evaluation: A traditional evaluation protocol involves selecting half of the test dataset's classes, repeating the process ten times, and reporting the mean accuracy with a standard deviation of ten times.

• Full-classes Evaluation: Perform evaluation on the entire dataset.

# On ActivityNet: reporting the half-classes and full-classes results
sh scripts/run_test_zeroshot.sh  configs/anet/anet_zero_shot.yaml exp/k400/ViT-L/14/f8/last_model.pt

# On UCF101: reporting the half-classes and full-classes results
sh scripts/run_test_zeroshot.sh  configs/ucf101/ucf_zero_shot.yaml exp/k400/ViT-L/14/f8/last_model.pt

# On HMDB51: reporting the half-classes and full-classes results
sh scripts/run_test_zeroshot.sh  configs/hmdb51/hmdb_zero_shot.yaml exp/k400/ViT-L/14/f8/last_model.pt

# On Kinetics-600: manually calculating the mean accuracy with standard deviation of three splits.
sh scripts/run_test.sh  configs/k600/k600_zero_shot_split1.yaml exp/k400/ViT-L/14/f8/last_model.pt
sh scripts/run_test.sh  configs/k600/k600_zero_shot_split2.yaml exp/k400/ViT-L/14/f8/last_model.pt
sh scripts/run_test.sh  configs/k600/k600_zero_shot_split3.yaml exp/k400/ViT-L/14/f8/last_model.pt

📌 BibTeX & Citation

If you find this repository useful, please star🌟 this repo and cite📑 our paper:

@inproceedings{wu2023revisiting,
  title={Revisiting classifier: Transferring vision-language models for video recognition},
  author={Wu, Wenhao and Sun, Zhun and Ouyang, Wanli},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={37},
  number={3},
  pages={2847--2855},
  year={2023}
}

@article{wu2023transferring,
  title={Transferring vision-language models for visual recognition: A classifier perspective},
  author={Wu, Wenhao and Sun, Zhun and Song, Yuxin and Wang, Jingdong and Ouyang, Wanli},
  journal={International Journal of Computer Vision},
  pages={1--18},
  year={2023},
  publisher={Springer}
}

If you also find BIKE useful, please cite the paper:

@inproceedings{bike,
  title={Bidirectional Cross-Modal Knowledge Exploration for Video Recognition with Pre-trained Vision-Language Models},
  author={Wu, Wenhao and Wang, Xiaohan and Luo, Haipeng and Wang, Jingdong and Yang, Yi and Ouyang, Wanli},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023}
}

🎗️ Acknowledgement

This repository is built based on ActionCLIP and CLIP. Sincere thanks to their wonderful works.

👫 Contact

For any question, please file an issue.