💥 November 18th 2021: The code of Long Short-Term Transformer (LSTR) is released here.

💥 July 08th 2021: We are releasing Long Short-Term Transformer (LSTR), a more effective and efficient method for modeling prolonged sequence data! LSTR achieves SoTA on Online Action Detection benchmarks.

💥 May 25th 2021: For future comparison with TRN using Kinetics pretrained features, we report our results on THUMOS as 62.1% in mAP, on TVSeries as 86.2% in cAP, and on HACS Segment as 78.9% in mAP.

For feature encoding, we use ResNet-50 model for the RGB input, and the BN-Inception model for the optical flow input. To replicate our results, please use the pretrained weights of ResNet-50 in MMAction2 and BN-Inception in this repo.

This is a PyTorch reimplementation for our ICCV 2019 paper "Temporal Recurrent Networks for Online Action Detection".

• The code is developed with CUDA 9.0, Python >= 3.6, PyTorch >= 1.0

1. Download the HDD and THUMOS'14 datasets.

2. Extract feature representations for video frames.

   • For HDD dataset, we use the Inception-ResNet-V2 pretrained on ImageNet for the RGB input.

   • For THUMOS'14 dataset, we use the ResNet-200 model for the RGB input, and the BN-Inception model for the optical flow input. To replicate our results, please follow the repo here: https://github.com/yjxiong/anet2016-cuhk.

   Note: We compute the optical flow for the THUMOS'14 dataset using FlowNet2.0.

3. If you want to use our dataloaders, please make sure to put the files as the following structure:

   • HDD dataset:

   $YOUR_PATH_TO_HDD_DATASET
   ├── inceptionresnetv2/
   |   ├── 201702271017.npy (of size L x 1536 x 8 x 8)
   │   ├── ...
   ├── sensor/
   |   ├── 201702271017.npy (of size L x 8)
   |   ├── ...
   ├── target/
   |   ├── 201702271017.npy (of size L)
   |   ├── ...
   

   • THUMOS'14 dataset:

   $YOUR_PATH_TO_THUMOS_DATASET
   ├── resnet200-fc/
   |   ├── video_validation_0000051.npy (of size L x 2048)
   │   ├── ...
   ├── bn_inception/
   |   ├── video_validation_0000051.npy (of size L x 1024)
   |   ├── ...
   ├── target/
   |   ├── video_validation_0000051.npy (of size L x 22)
   |   ├── ...
   

You can skip the step of 1, 2, 3 above and directly use the pre-extracted features and targets from TeSTra. They extactly follow our data structure and should be able to reproduce TRN's performance. However, if you have any question about the processing of these features and targets, please contact the authors of TeSTra directly.

4. Create softlinks of datasets:

   cd TRN.pytorch
   ln -s $YOUR_PATH_TO_HDD_DATASET data/HDD
   ln -s $YOUR_PATH_TO_THUMOS_DATASET data/THUMOS
   

• Single GPU training on HDD dataset:

cd TRN.pytorch
# Training from scratch
python tools/trn_hdd/train.py --gpu $CUDA_VISIBLE_DEVICES
# Finetuning from a pretrained model
python tools/trn_hdd/train.py --checkpoint $PATH_TO_CHECKPOINT --gpu $CUDA_VISIBLE_DEVICES

• Multi-GPU training on HDD dataset:

cd TRN.pytorch
# Training from scratch
python tools/trn_hdd/train.py --gpu $CUDA_VISIBLE_DEVICES --distributed
# Finetuning from a pretrained model
python tools/trn_hdd/train.py --checkpoint $PATH_TO_CHECKPOINT --gpu $CUDA_VISIBLE_DEVICES --distributed

• Single GPU training on THUMOS'14 dataset:

cd TRN.pytorch
# Training from scratch
python tools/trn_thumos/train.py --gpu $CUDA_VISIBLE_DEVICES
# Finetuning from a pretrained model
python tools/trn_thumos/train.py --checkpoint $PATH_TO_CHECKPOINT --gpu $CUDA_VISIBLE_DEVICES

• Multi-GPU training on THUMOS'14 dataset:

cd TRN.pytorch
# Training from scratch
python tools/trn_thumos/train.py --gpu $CUDA_VISIBLE_DEVICES --distributed
# Finetuning from a pretrained model
python tools/trn_thumos/train.py --checkpoint $PATH_TO_CHECKPOINT --gpu $CUDA_VISIBLE_DEVICES --distributed

• HDD dataset:

cd TRN.pytorch
python tools/trn_hdd/eval.py --checkpoint $PATH_TO_CHECKPOINT --gpu $CUDA_VISIBLE_DEVICES

• THUMOS'14 dataset:

cd TRN.pytorch
python tools/trn_thumos/eval.py --checkpoint $PATH_TO_CHECKPOINT --gpu $CUDA_VISIBLE_DEVICES

NOTE: There are two kinds of evaluation methods in our code. (1) Using --debug during training considers each short video clip (consisting of 90 and 64 consecutive frames for HDD and THUMOS'14 datasets, respectively) as one test sample, and separately runs inference and evaluates on all short video clips (even though some of them are from the same long video). (2) Using eval.py after training runs inference and evaluates on long videos (frame by frame, from the beginning to the end), which is the evaluation method we reported in the paper.

If you are using the data/code/model provided here in a publication, please cite our paper:

@inproceedings{onlineaction2019iccv,
    title = {Temporal Recurrent Networks for Online Action Detection},
    author = {Mingze Xu and Mingfei Gao and Yi-Ting Chen and Larry S. Davis and David J. Crandall},
    booktitle = {IEEE International Conference on Computer Vision (ICCV)},
    year = {2019}
}