Back To Reality: Weak-supervised 3D Object Detection with Shape-guided Label Enhancement

Announcement 🔥

We have not tested the code yet. We will finish this project by April.

Introduction

This repo contains PyTorch implementation for paper Back To Reality: Weak-supervised 3D Object Detection with Shape-guided Label Enhancement (CVPR2022)

@inproceedings{xu2022br,
author = {Xu, Xiuwei and Wang, Yifan and Zheng, Yu and Rao, Yongming and Lu, Jiwen and Zhou, Jie},
title = {Back To Reality: Weak-supervised 3D Object Detection with Shape-guided Label Enhancement},
booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2022}
}

Other papers related to 3D object detection with synthetic shape:

RandomRooms: Unsupervised Pre-training from Synthetic Shapes and Randomized Layouts for 3D Object Detection (ICCV 2021)

New dataset 💥

We conduct additional experiment on the more challenging Matterport3D dataset. From ModelNet40 and Matterport3D, we select all 13 shared categories, each containing more than 80 object instances in Matterport3D training set, to construct our benchmark (Matterport3d-md40). Below is the performance of FSB, WSB and BR (point-version) based on Votenet:

Note that we use OpenCV to estimate the rotated bounding boxes (RBB) as ground-truth, instead of the axis-aligned bounding boxes used in ScanNet-md40 benchmark.

ScanNet-md40 and Matterport3d-md40 are two more challenging benckmarks for indoor 3D object detection. We hope they will promote future research on small object detection and synthetic-to-real scene understanding.

Dependencies

We evaluate this code with Pytorch 1.8.1 (cuda11), which is based on the official implementation of Votenet and GroupFree3D. Please follow the requirements of them to prepare the environment. Other packages can be installed using:

pip install open3d sklearn tqdm

Current code base is tested under following environment:

Python 3.6.13
PyTorch 1.8.1
numpy 1.19.2
open3d 0.12.0
opencv-python 4.5.1.48
plyfile 0.7.3
scikit-learn 0.24.1

Data preparation

ScanNet

To start from the raw data, you should:

Follow the README under GroupFree3D/scannet or Votenet/scannet to generate the real scenes.
Follow the README under ./data_generation/ScanNet to generate the virtual scenes.

The processed data can also be downloaded from here. They should be placed to paths:

./detection/Votenet/scannet/
./detection/GroupFree3D/scannet/

After that, the file directory should be like:

...
└── Votenet (or GroupFree3D)
    ├── ...
    └── scannet
        ├── ...
        ├── scannet_train_detection_data_md40
        ├── scannet_train_detection_data_md40_obj_aug
        └── scannet_train_detection_data_md40_obj_mesh_aug

Matterport3D

To start from the raw data, you should:

Follow the README under Votenet/scannet to generate the real scenes.
Follow the README under ./data_generation/Matterport3D to generate the virtual scenes.

The processed data can also be downloaded from here.

The file directory should be like:

...
└── Votenet
    ├── ...
    └── matterport
        ├── ...
        ├── matterport_train_detection_data_md40
        ├── matterport_train_detection_data_md40_obj_aug
        └── matterport_train_detection_data_md40_obj_mesh_aug

Usage

Please follow the instructions below to train different models on ScanNet. Change --dataset scannet to --dataset matterport for training on Matterport3D.

Votenet

1. Fully-Supervised Baseline

To train the Fully-Supervised Baseline (FSB) on Scannet data:

# Recommended GPU num: 1

cd Votenet

CUDA_VISIBLE_DEVICES=0 python train_Votenet_FSB.py --dataset scannet --log_dir log_Votenet_FSB --num_point 40000

2. Weakly-Supervised Baseline

To train the Weakly-Supervised Baseline (WSB) on Scannet data:

# Recommended num of GPUs: 1

CUDA_VISIBLE_DEVICES=0 python train_Votenet_WSB.py --dataset scannet --log_dir log_Votenet_WSB --num_point 40000

3. Back To Reality

To train BR (mesh-version) on Scannet data, please run:

# Recommended num of GPUs: 2

CUDA_VISIBLE_DEVICES=0,1 python train_Votenet_BR.py --dataset scannet --log_dir log_Votenet_BRM --num_point 40000

CUDA_VISIBLE_DEVICES=0,1 python train_Votenet_BR_CenterRefine --dataset scannet --log_dir log_Votenet_BRM_Refine --num_point 40000 --checkpoint_path log_Votenet_BRM/train_BR.tar

To train BR (point-version) on Scannet data, please run:

# Recommended num of GPUs: 2

CUDA_VISIBLE_DEVICES=0,1 python train_Votenet_BR.py --dataset scannet --log_dir log_Votenet_BRP --num_point 40000 --dataset_without_mesh

CUDA_VISIBLE_DEVICES=0,1 python train_Votenet_BR_CenterRefine --dataset scannet --log_dir log_Votenet_BRP_Refine --num_point 40000 --checkpoint_path log_Votenet_BRP/train_BR.tar --dataset_without_mesh

GroupFree3D

1. Fully-Supervised Baseline

To train the Fully-Supervised Baseline (FSB) on Scannet data:

# Recommended num of GPUs: 4

cd GroupFree3D

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node <num_of_gpus_to_use> train_GF_FSB.py --num_point 50000 --num_decoder_layers 6 --size_delta 0.111111111111 --center_delta 0.04 --learning_rate 0.006 --decoder_learning_rate 0.0006 --weight_decay 0.0005 --dataset scannet --log_dir log_GF_FSB --batch_size 4

2. Weakly-Supervised Baseline

To train the Weakly-Supervised Baseline (WSB) on Scannet data:

# Recommended num of GPUs: 4

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node <num_of_gpus_to_use> train_GF_WSB.py --num_point 50000 --num_decoder_layers 6 --size_delta 0.111111111111 --center_delta 0.04 --learning_rate 0.006 --decoder_learning_rate 0.0006 --weight_decay 0.0005 --dataset scannet --log_dir log_GF_WSB --batch_size 4

3. Back To Reality

To train BR (mesh-version) on Scannet data, please run:

# Recommended num of GPUs: 4

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node <num_of_gpus_to_use> train_GF_BR.py --num_point 50000 --num_decoder_layers 6 --size_delta 0.111111111111 --center_delta 0.04 --learning_rate 0.006 --decoder_learning_rate 0.0006 --weight_decay 0.0005 --dataset scannet --log_dir log_GF_BRM --batch_size 4

# Recommended num of GPUs: 6

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node <num_of_gpus_to_use> train_GF_BR_CenterRefine.py --num_point 50000 --num_decoder_layers 6 --size_delta 0.111111111111 --center_delta 0.04 --learning_rate 0.001 --decoder_learning_rate 0.0006 --weight_decay 0.0005 --dataset scannet --log_dir log_GF_BRM_Refine --checkpoint_path <[checkpoint_path_of_groupfree3D]/ckpt_epoch_last.pth> --max_epoch 120 --val_freq 10 --save_freq 20 --batch_size 2

To train BR (point-version) on Scannet data, please run:

# Recommended num of GPUs: 4

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node <num_of_gpus_to_use> train_GF_BR.py --num_point 50000 --num_decoder_layers 6 --size_delta 0.111111111111 --center_delta 0.04 --learning_rate 0.006 --decoder_learning_rate 0.0006 --weight_decay 0.0005 --dataset scannet --log_dir log_GF_BRP --batch_size 4 --dataset_without_mesh

# Recommended num of GPUs: 6

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node <num_of_gpus_to_use> train_GF_BR_CenterRefine.py --num_point 50000 --num_decoder_layers 6 --size_delta 0.111111111111 --center_delta 0.04 --learning_rate 0.001 --decoder_learning_rate 0.0006 --weight_decay 0.0005 --dataset scannet --log_dir log_GF_BRP_Refine --checkpoint_path <[checkpoint_path_of_groupfree3D]/ckpt_epoch_last.pth> --max_epoch 120 --val_freq 10 --save_freq 20 --batch_size 2 --dataset_without_mesh

TODO list

We will add the following to this repo:

Virtual scene generation for Matterport3D
Data and code for training Votenet (both baseline and BR) on the Matterport3D dataset

Acknowledgements

We thank a lot for the flexible codebase of Votenet and GroupFree3D.

[CVPR 2022] Back To Reality: Weak-supervised 3D Object Detection with Shape-guided Label Enhancement

Related tags

Overview

Back To Reality: Weak-supervised 3D Object Detection with Shape-guided Label Enhancement

Announcement 🔥

Introduction

New dataset 💥

Dependencies

Data preparation

ScanNet

Matterport3D

Usage

Votenet

1. Fully-Supervised Baseline

2. Weakly-Supervised Baseline

3. Back To Reality

GroupFree3D

1. Fully-Supervised Baseline

2. Weakly-Supervised Baseline

3. Back To Reality

TODO list

Acknowledgements

Owner

Xiuwei Xu

PyTorch-lightning implementation of the ESFW module proposed in our paper Edge-Selective Feature Weaving for Point Cloud Matching

This is the implementation of "SELF SUPERVISED REPRESENTATION LEARNING WITH DEEP CLUSTERING FOR ACOUSTIC UNIT DISCOVERY FROM RAW SPEECH" submitted to ICASSP 2022

Swin-Transformer is basically a hierarchical Transformer whose representation is computed with shifted windows.

Source code of our BMVC 2021 paper: AniFormer: Data-driven 3D Animation with Transformer

Source code for "Roto-translated Local Coordinate Framesfor Interacting Dynamical Systems"

Voxel-based Network for Shape Completion by Leveraging Edge Generation (ICCV 2021, oral)

SenseNet is a sensorimotor and touch simulator for deep reinforcement learning research

A PyTorch implementation of EventProp [https://arxiv.org/abs/2009.08378], a method to train Spiking Neural Networks

ColBERT: Contextualized Late Interaction over BERT (SIGIR'20)

Simple PyTorch implementations of Badnets on MNIST and CIFAR10.

Implementation of "Fast and Flexible Temporal Point Processes with Triangular Maps" (Oral @ NeurIPS 2020)

A Strong Baseline for Image Semantic Segmentation

Pytorch implementation of U-Net, R2U-Net, Attention U-Net, and Attention R2U-Net.

Code and experiments for "Deep Neural Networks for Rank Consistent Ordinal Regression based on Conditional Probabilities"

Pytorch-Swin-Unet-V2 - a modified version of Swin Unet based on Swin Transfomer V2

The world's simplest facial recognition api for Python and the command line

InvTorch: memory-efficient models with invertible functions

Repository for the paper "Exploring the Sensory Spaces of English Perceptual Verbs in Natural Language Data"

Colour detection is necessary to recognize objects, it is also used as a tool in various image editing and drawing apps.

Experiments for distributed optimization algorithms