Awesome

RTM3D: Real-time Monocular 3D Detection from Object Keypoints for Autonomous Driving

Monocular 3D Detection with Geometric Constraints Embedding and Semi-supervised Training (KM3D)

RTM3D(ECCV2020) and KM3D (namely RTM3D++) are efficiency and accuracy monocular 3D object detection methods for autonomous driving.

We replaced the post-processing of RTM3D with KM3D's Geometric Reasoning Module (GRM) to increase the speed of inference. KM3D, RTM3D

Introduction

RTM3D is a novel one-stage and keypoints-based framework for monocular 3D objects detection. RTM3D is the first real-time system (FPS>24) for monocular image 3D detection while achieves state-of-the-art performance on the KITTI benchmark. KM3D reformulate the geometric constraints as a differentiable version and embed it into the net-work to reduce running time while maintaining the consistency of model outputs in an end-to-end fashion. KM3D achieves 46FPS and SOTA performance on the KITTI benchmark. RTM3D and KM3D only require RGB images without synthetic data, instance segmentation, CAD model, or depth generator.

Highlights

• Fast: 47FPS of single image test speed in KITTI benchmark with 384*1280 resolution
• Accuracy: SOTA on the KITTI benchmark.
• Anchor Free: No 2D or 3D anchor are reauired
• Differentiable geometric reasoning module: Promote the running efficiency and optimize outputs of network jointly. Combining the strengths of both CNN and geometric constraints.
• Easy to deploy: RTM3D and KM3D only uses conventional convolution and upsampling operations, and the geometry module only needs to solve SVD, so it is very easy to deploy and accelerate.

KM3D Baseline and Model Zoo

All experiments are tested with Ubuntu 16.04, Pytorch 1.0.0, CUDA 9.0, Python 3.6, single NVIDIA 1080Ti

IoU Setting 1: Car IoU > 0.5, Pedestrian IoU > 0.25, Cyclist IoU > 0.25

IoU Setting 2: Car IoU > 0.7, Pedestrian IoU > 0.5, Cyclist IoU > 0.5

• Training on KITTI train split and evaluation on val split.
  • Backbone: ResNet-18
  • FPS: 46.7
  • Model: (Google Drive), (Baidu Cloud 提取码：60ks)

• Training on KITTI train split and evaluation on val split.
  • Backbone: DLA-34
  • FPS: 28.6
  • Model: (Google Drive), (Baidu Cloud 提取码：1h6s)

• Training on KITTI train split with right images augmentation and evaluation on val split.
  • Backbone: ResNet-18
  • FPS: 46.7
  • Model: (Google Drive), (Baidu Cloud 提取码：sr23)

• Training on KITTI train split with right images augmentation and evaluation on val split.
  • Backbone: DLA-34
  • FPS: 28.6
  • Model: (Google Drive), (Baidu Cloud 提取码：qqk6)

Installation

Please refer to INSTALL.md

Dataset preparation

Please download the official KITTI 3D object detection dataset and organize the downloaded files as follows:

KM3DNet
├── kitti_format
│   ├── data
│   │   ├── kitti
│   │   |   ├── annotations 
│   │   │   ├── calib /000000.txt .....
│   │   │   ├── image(left[0-7480] right[7481-14961] input augmentatiom)
│   │   │   ├── label /000000.txt .....
|   |   |   ├── train.txt val.txt trainval.txt
├── src
├── demo_kitti_format
├── readme
├── requirements.txt

Quick Demo

Please refer to DEMO.md for a quick demo to test with a pretrained model and visualize the predicted results on your custom data or the original KITTI data.

Getting Started

Please refer to GETTING_STARTED.md to learn more usage about this project.

Acknowledgement

• CenterNet

License

RTM3D and KM3D are released under the MIT License (refer to the LICENSE file for details). Portions of the code are borrowed from, CenterNet, dla (DLA network), DCNv2(deformable convolutions), iou3d and kitti_eval (KITTI dataset evaluation). Please refer to the original License of these projects (See NOTICE).

Citation

If you find this project useful for your research, please use the following BibTeX entry.

@misc{2009.00764,
Author = {Peixuan Li},
Title = {Monocular 3D Detection with Geometric Constraints Embedding and Semi-supervised Training},
Year = {2020},
Eprint = {arXiv:2009.00764},
}
@misc{2001.03343,
Author = {Peixuan Li and Huaici Zhao and Pengfei Liu and Feidao Cao},
Title = {RTM3D: Real-time Monocular 3D Detection from Object Keypoints for Autonomous Driving},
Year = {2020},
Eprint = {arXiv:2001.03343},
}