Network Grafting Algorithm

Code release for "Learning to Exploit Multiple Vision Modalities by Using Grafted Networks", ECCV 2020.

If you use this project, please cite:

• Y. Hu, T. Delbruck, S-C. Liu, "Learning to Exploit Multiple Vision Modalities by Using Grafted Networks" in The 16th European Conference on Computer Vision (ECCV), Online, 2020.

Install evtrans locally

python setup.py develop

Project structure

1. evtrans folder contains utility scripts and network definitions. Make sure you run setup.py to install this module locally.

2. configs folder contains pretrained network configurations and class definitions.

3. scripts folder contains:

   • nmnist: scripts that are related to N-MNIST classification.
   • object_detection: inference scripts related to thermal and event data experiments.
   • prepare-data: how to prepare data for N-MNIST, MVSEC and FLIR datasets.
   • weights: you should put the weights here. Small weight files are uploaded in this folder and a Google drive shared folder (see following).

Run the script

1. Object Detection on Thermal Driving Dataset

   python export_thermal_yolo_results.py --val_data_dir /path/to/val/data --checkpoint /path/to/checkpoint.pt --detection_path /path/to/dump/prediction/result --conv_input_dim 1 --img_size 640 --cut_stage [1, 2, or 3] [--vis]
   

2. Car Detection on Event Camera Driving Dataset

   python export_ev_yolo_results.py --img_size 346 --val_data_dir /path/to/val/data --checkpoint /path/to/checkpoint.pt --detection_path /path/to/dump/prediction/result --conv_input_dim [3 or 10] --cut_stage [1, 2 or 3]
   

3. N-MNIST Classification

   python val_pt_nmnist.py --test_path /path/to/val/data --checkpoint /path/to/checkpoint.pt
   

Trained weights

Some selected models are shared here.

This folder contains the original YOLOv3 pretrained weights and trained GN frontend for each task and configuration.

Data

Data for evaluation is released here.

Alternative Downloads

If you cannot access above links, please go to Zenodo platform and download, use this link

Contacts

Yuhuang Hu
[email protected]