Here is the pytorch code for our paper NODIS: Neural Ordinary Differential Scene Understanding (ECCV 2020). If the paper is helpful for you, we request that you cite our work.

• Short Video:https://youtu.be/4VLnOpeIzjs
• Long Video: https://youtu.be/kgMRG8LxkH0

Our code is supported by neural-motifs and torchdiffeq. Great thanks to Rowan Zellers and Ricky Chen!

1. Install python and pytorch if you haven't. Our code is based on python 3.6 and pytorch 0.4.1.

2. Compile: run make in the main directory

3. Download Neural ODE module here

4. For a fair comparison we use the pretrained object detector checkpoint provided by neural-motifs. You can download it here directly and save it under checkpoints/vgdet/

5. The final directories for data and detection models should look like:

|-- checkpoints
|   |-- vgdet
|-- data
|   |-- stanford_filtered
|-- dataloaders
|-- lib
|-- torchdiffeq
|-- models

You can train the NODIS model with train_rel.py. We trained the model on a GTX 1080Ti.

• For PredCLS:

python train_rels.py -m predcls -order random -b 6 -p 100 -lr 1e-4 -ngpu 1 -ckpt checkpoints/vgdet/vg-24.tar -save_dir checkpoints/ -nepoch 20

• For SGCLS:

python train_rels.py -m sgcls -order random -b 6 -p 100 -lr 1e-4 -ngpu 1 -ckpt checkpoints/vgdet/vg-24.tar -save_dir checkpoints/ -nepoch 20

• For SGGEN:

python train_rels.py -m sgdet -order random -b 6 -p 100 -lr 1e-4 -ngpu 1 -ckpt $CHECKPOINT -save_dir checkpoints/ -nepoch 20

You can evaluate the model trained by yourself.

• For PredCLS:

python eval_rels.py -m predcls -order random -b 6 -p 100 -lr 1e-3 -ngpu 1 -test -ckpt $CHECKPOINT -nepoch 50

• For SGCLS:

python eval_rels.py -m sgcls -order random -b 6 -p 100 -lr 1e-3 -ngpu 1 -test -ckpt $CHECKPOINT -nepoch 50

• For SGGEN:

python eval_rels.py -m sgdet -order random -b 6 -p 100 -lr 1e-3 -ngpu 1 -test -ckpt $CHECKPOINT -nepoch 50

or you can download the pretrained NODIS PREDCLS/SGCLS/SGGEN here.

This is a draft version, if you find any problem, please contact with us.