Awesome

Video based Object Pose Estimation using Transformers

This directory contains implementation of paper Video based Object 6D Pose Estimation using Transformers.
Accepted into NeuRIPS 2022 Workshop on Vision Transformers: Theory and Applications.

If this code helps with your work, please cite:

@article{beedu2022video,
  title={Video based Object 6D Pose Estimation using Transformers},
  author={Beedu, Apoorva and Alamri, Huda and Essa, Irfan},
  journal={arXiv preprint arXiv:2210.13540},
  year={2022}
}

Environment setup

Please install all the requirements using requirements.txt pip3 install -r requirements.txt

Directory setup

Create a ./evaluation_results_video, wandb, logs, output and model folders.

Arguments

Arguments and their defaults are in arguments.py

• backbone swin or beit
• use_depth To use ground-truth depth during training
• restore_file name of the file in --model_dir_path containing weights to reload before training
• lr Learning rate for the optimiser
• batch_size Batch size for the dataset
• workers num_workers
• env_name environment name for wandb, which is also the checkpoint name

Setting up dataset

Download the entire YCB dataset from https://rse-lab.cs.washington.edu/projects/posecnn/
The data folder looks like

train_eval.py
dataloader.py
├── data
│   ├── YCB
│   │   └── data
│   │       ├── 0000
│   │       └── 0001
│   │   └── models
│   │   └── train.txt
│   │   └── keyframe.txt
│   │   └── val.txt

Execution

python3 train_eval.py --batch_size=8 --lr=0.0001 --backbone=swin --predict_future=1 --use_depth=1 --video_length=5 --workers=12

Video based Object Pose Estimation using Transformers

This directory contains implementation for estimating 6D object poses from videos.

Environment setup

Please install all the requirements using requirements.txt pip3 install -r requirements.txt

Directory setup

Create a ./evaluation_results_video, wandb, logs, output and model folders.

Arguments

Arguments and their defaults are in arguments.py

• backbone swin or beit
• use_depth To use ground-truth depth during training
• restore_file name of the file in --model_dir_path containing weights to reload before training
• lr Learning rate for the optimiser
• batch_size Batch size for the dataset
• workers num_workers
• env_name environment name for wandb, which is also the checkpoint name

Setting up dataset

Download the entire YCB dataset from https://rse-lab.cs.washington.edu/projects/posecnn/

Download the checkpoint from https://drive.google.com/drive/folders/1lQh3G7KN-SHb7B-NYpqWj55O1WD4E9s6?usp=sharing

Add the checkpoint to ./model/Videopose/last_checkpoint_0000.pt, and pass the argument --restore_file=Videopose during training to start from a checkpoint. If no start_epoch is mentioned, the training will restart from the last checkpoint.

The data folder looks like

train_eval.py
dataloader.py
├── data
│   ├── YCB
│   │   └── data
│   │       ├── 0000
│   │       └── 0001
│   │   └── models
│   │   └── train.txt
│   │   └── keyframe.txt
│   │   └── val.txt

Execution

The project uses wandb for visualisation.

Main branch uses -posecnn.mat files, that I manually generated for every frame in the dataset using Posecnn repository. If you do not have those files, v1 is the branch to use.

python3 train_eval.py --batch_size=8 --lr=0.0001 --backbone=swin --predict_future=1 --use_depth=1 --video_length=5 --workers=12

Evaluation

Evaluation currently runs only on one GPU.

python3 train_eval.py --batch_size=8 --backbone=swin --predict_future=1 --use_depth=1 --video_length=5 --workers=12  --restore_file=Videopose --split=eval

The command will create several mat files for the keyframes and also saves images into a folder. To evaluate the mat files, please use the YCBToolBox.