ANCSH_Pytorch (Category Level Articulated Object Pose Estimation Pytorch)

Overview

This repository contains the implementation of the paper Category-Level Articulated Object Pose Estimation
Checkout the official code release for the paper at dragonlong/articulated-pose..

Paper Website

Content

Additional Repo
Setup
Dataset
Pretrained Models
Training
Inference & Optimization
Evaluation & Visualization

Additional-Repo

OPD: In this paper, we reimplement ANCSH using PyTorch for a comparison. For details, check the OPD website.

Setup

The implementation has been tested on Ubuntu 20.04, with PyTorch 1.7.1, CUDA 11 and CUDNN 8.0.3.

Clone the repository

git clone git@github.com:3dlg-hcvc/ANCSH-pytorch.git

Setup python environment

conda create -n ancsh python=3.7 
conda activate ancsh  
pip install -e .
pip install -r requirements.txt
pip install "git+git://github.com/erikwijmans/Pointnet2_PyTorch.git#egg=pointnet2_ops&subdirectory=pointnet2_ops_lib"

(If meeting unauthenticated git protocol on port 9418 ), run git config --global url."https://".insteadOf git:// for above last command to install pointnet2_ops_lib)

Dataset

You can download the original processed [eyeglasses] (eyeglasses_ancsh) and our [OneDoor] (dataset) dataset for ANCSH to ./dataset folder (there is one dataset.tar.gz under /dataset/ANCSH).

Pretrained-Models

You can download our pretrained models to ./models folder (there is one models.tar.gz under /models/ANCSH).

[eyeglasses-ancsh] [eyeglasses-npcs]

[onedoor-ancsh] [onedoor-npcs]

Training

To train from the scratch, you can use below commands. For this paper, we need to train two independent models (ANCSH, NPCS), and combine their results as final results. (We have got the author's reply on the details)

We use hydra in this project, please modify the config files in configs/paths to change the path setting (Refer to cc_ancsh.yaml and cc_npcs.yaml).

Train the two models
```
python train.py \
network=<NETWORK_TYPE> \
paths=<PATH_CONFIG_OPTION> \
train.input_data=<PATH_TO_TRAIN_DATA> \
test.input_data=<PATH_TO_TEST_DATA> 
```
- Network Type:
  - ANCSH Model:
    - network=ancsh
  - NPCS Model:
    - network=npcs
- Dataset:
  - Eyeglasses: (The same split to original ANCSH paper)
    - train.input_data=<PATH_TO_REPO>/dataset/eyeglasses_ancsh/train.h5
    - test.input_data=<PATH_TO_REPO>/dataset/eyeglasses_ancsh/test.h5
  - OneDoor:
    - train.input_data=<PATH_TO_REPO>/dataset/OneDoor/train.h5
    - test.input_data=<PATH_TO_REPO>/dataset/OneDoor/val.h5

Inference&Optimization

Inference with pretrained Model

python train.py \
network=<NETWORK_TYPE> \
paths=<PATH_CONFIG_OPTION> \
train.input_data=<PATH_TO_TRAIN_DATA> \
test.input_data=<PATH_TO_TEST_DATA> 
eval_only=true \
test.inference_model=<PRETRAINED_MODEL>

The options are the same to above.

Optimization
```
python optimize.py \
paths=<PATH_CONFIG_OPTION> \
ancsh_results_path=<ANCSH_RESULTS> \
npcs_results_path=<NPCS_RESULTS> \
num_parts=<NUM_PARTS>
```
The options are the same to above. The results file are from the inference of above two models.
- Dataset:
  - Eyeglasses: (The same split to original ANCSH paper)
    - num_parts=3
  - OneDoor:
    - num_parts=2

Evaluation with original ANCSH evaluation metric

python evaluate.py \
paths=<PATH_CONFIG_OPTION> \
optimization_result_path=<PATH_TO_OPTIMIZATION_RESULT> \
num_parts=<NUM_PARTS>

The options are the same to above.

Evaluation&Visualization

You can ignore this section if you want to use ANCSH directly without considering OPD.

We convert the results of OneDoor dataset into OPD inference results format, evalaute and visualize using the code from OPD

Get the prediction.h5 from above evaluation, and conver the results into OPDRCNN format (Use the code from OPD). Need to download the dataset for OPDRCNN in OPD.)

python <PATH_TO_OPD>/ANCSH_convert_results.py \
--ancsh-result-path <PATH_TO_FINAL_RESULT> \
--data-path <PATH_TO_DATASET>
* Dataset:
  * OneDoor Data:
    * --data-path <PATH_TO_REPO>/dataset/MotionDataset_h5_ancsh

After converting, get the inference file in the OPD format, check Evaluation in README of OPD. (Use --inference-file option, change the dataset to MotionDataset_h5_ancsh)