Dual Transfer Learning for Event-based End-task Prediction via Pluggable Event to Image Translation (ICCV'21)
We have updated our paper by correcting some typos and adding more references.
Please refer to the Arxiv paper at https://arxiv.org/pdf/2109.01801.pdf for the latest information.
If you find this resource helpful, please cite the paper as follows:
@inproceedings{wang2021dual,
title={Dual transfer learning for event-based end-task prediction via pluggable event to image translation},
author={Wang, Lin and Chae, Yujeong and Yoon, Kuk-Jin},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={2135--2145},
year={2021}
}
- Segmentic segmentation on the general visual condition.
- Segmentic segmentation on over-exposed visual condition.
- Depth estimation on the night-scene condition
Download
git clone https://github.com/addisonwang2013/DTL/
Make your own environment
conda create -n myenv python=3.7
conda activate myenv
Install the requirements
cd evdistill
pip install -r requirements.txt
Download example validation data (general and LDR visual condtions) from this link: DDD17 example data
- For DDD17 dataset general visiual condition, please put the dataset to
./dataset/ddd17/general/
- For DDD17 dataset low dynamic range (LDR) condition, please put the dataset to
./dataset/ddd17/ldr/
Download the pretrained models from this link: checkpoints
- Put the checkpoint of event segmentation network into
./res/
Modify the python configurations.py
in the configs
folder with the relevant paths to the test data and checkpoints
- For the test data, e.g. DDD17, please assign the path to
./test_dir/ddd17/
- For the checkpoint of event network, please assign the path to
./res/eventdual_best.pth
Visualizing semantic segmentation results for general and LDR visual condtions:
python visualize.py
In this work, for convenience, the event data are embedded and stored as multi-channel event images, which are the paired with the aps frames. It is also possible to directly feed event raw data after embedding to the student network directly with aps frames.
The skeleton code is inspired by Deeplab-v3-Plus and EDSR (https://github.com/sanghyun-son/EDSR-PyTorch)