Fuzzy-NMS

Project purpose

Implementation of the paper "Fuzzy-NMS: Improving 3D Object Detection with Fuzzy Classification in NMS" on Openpcdet.

Installation

Requirements

All the codes are tested in the following environment:

• Linux (tested on Ubuntu 14.04/16.04/18.04/20.04/21.04)
• Python 3.6+
• PyTorch 1.1 or higher (tested on PyTorch 1.1, 1,3, 1,5~1.10)
• CUDA 9.0 or higher (PyTorch 1.3+ needs CUDA 9.2+)
• spconv v1.0 (commit 8da6f96) or spconv v1.2 or spconv v2.x

Please refer to the official installation steps of Openpcdet.

Introduction to Core Documentation

pcdet
├── datasets
├── models
│   ├── backbones_2d
│   ├── backbones_3d
│   ├── dense_heads
│   ├── detectors
│		├── detector3d_template.py #A part is added in lines 283 to 294, and the dictionary composed of nms  │                                   results is counted, and a return value nms_dicts is added to the          │                 ┇                 post_processing function.
│		         
│		└── pointpillar.py #A return value nms_dicts is added to the forward function, and the same is true │                           for other baseline detector files.		          
│   ├── model_utils
│		├── fuzzy_code
│       	├── cpp_fuzzy.py #Loads a dynamic link library to classify objects by size and density.
│       	├── DBSCAN_plot.py #Plot the DBSCAN process.
│       	├── libfuzzy.so #C++ dynamic link library for fuzzy classification.
│		├── fuzzy_nms
│       	├── model_nms_utils_cpp.py #If it is a single run, you need to replace the code in this file 	│										with model_nms_utils.py in the upper directory.
│		├── model_nms_utils.py #The class_agnostic_nms function has been rewritten and a return value has   │                               been added.
├── ops
│   ├── iou3d_nms 
│		└──iou3d_nms_utils.py #Added soft-nms and Diou-nms.
tools
├── eval_utils
│   ├── eval_utils_fuzzy.py #It needs to be used when traversing to find the optimal parameters.
├── test.py #The eval_single_ckpt function needs to be modified when selecting a single run or a traversal    │            parameter.

Getting Started

Please refer to the official documents of Openpcdet to prepare the Kitti and Waymo datasets.

Test

cd tools
CUDA_VISIBLE_DEVICES=1 python test.py --cfg_file cfgs/kitti_models/pointpillar.yaml --batch_size 4 --ckpt pointpillar.pth --extra_tag nms_test

Results

KITTI 3D Object Detection Baselines

Selected supported methods are shown in the below table. The results are the 3D detection performance of moderate difficulty on the test set of KITTI dataset.

	Car	Pedestrian	Cyclist
PointPillars	73.13	34.16	54.68
PV-RCNN	78.71	40.01	62.05
IA-SSD	78.93	39.95	60.28
GD-MAE	76.03	36.46	54.99
BiProDet	81.77	45.71	64.01

KITTI BEV Object Detection Baselines

Selected supported methods are shown in the below table. The results are the BEV detection performance of moderate difficulty on the test set of KITTI dataset.

	Car	Pedestrian	Cyclist
PointPillars	86.79	39.74	62.20
PV-RCNN	87.71	45.37	66.12
IA-SSD	89.01	45.68	67.72
GD-MAE	88.38	41.32	62.46
BiProDet	89.02	50.97	70.34

Waymo Open Dataset Baselines

The following three baselines are all tested on val set of KITTI dataset.

	Vec_L1	Vec_L2	Ped_L1	Ped_L2	Cyc_L1	Cyc_L2
PointPillars	70.15/69.53	62.05/61.50	67.61/46.87	60.31/41.74	57.78/54.02	56.19/52.53
SECOND	69.07/68.44	60.85/60.29	65.15/53.73	58.29/47.98	53.72/52.24	52.35/50.91
M3DETR	76.58/75.92	69.24/68.60	67.10/57.73	59.00/50.68	68.12/66.72	66.46/65.09