Research-on-Pose-Estimation-of-Objects-1

物体位姿估计研究-1：复现位姿估计论文《Estimating 6D Pose From Localizing Designated Surface Keypoints》

物体位姿估计交流群：756535271，欢迎各位大佬入群交流！

Pipeline

Prepare Data

Go to folder data/ and follow its instructions. After this step, data will be prepared for training.

Keypoints Designation

Go to folder gendata/ and follow its instructions. After this step, you'll get models and keypoints raw data (in .ply format).

Object Detector

Go to folder detect/ and follow its instructions. After this step, you'll get YOLOv3 pre-trained weights for LINEMOD.

Keypoint Localization

Go to folder keypoint/ and follow its instructions. After this step, you'll get SPPE pre-trained weights for LINEMOD.

Pose Estimation

Run scripts/pose.sh to estimate the pose. Results will be saved to results/.

Evaluation

Run scripts/eval.sh to estimate the pose and evaluate the result

Results

ADD (-S) accuracy listed below. Eggbox and glue is calculated with ADD(-S).

Sequence	17 SIFT	17 Cluster	17 Cluster w/o dpg	9 Corner	9 SIFT
01 Ape	0.487	0.649	0.203	0.119	0.446
02 Benchvise	0.982	0.986	0.601	0.179	0.858
04 Camera	0.687	0.888	0.739	0.538	0.778
05 Can	0.821	0.921	0.873	0.231	0.813
06 Cat	0.645	0.863	0.661	0.318	0.632
08 Driller	0.803	0.961	0.936	0.217	0.596
09 Duck	0.508	0.587	0.523	0.100	0.329
10 Eggbox		0.977	0.924	0.971	0.305
11 Glue		0.954	0.519	0.667	0.810
12 Holepuncher	0.648	0.765	0.756	0.690	0.461
13 Iron	0.918	0.969	0.859	0.460	0.838
14 Lamp		0.973	0.904	0.675	0.961
15 Phone	0.730	0.865	0.823	0.510	0.786
Average		0.874	0.717	0.437	0.663

Occlusion LINEMOD

Sequence	Mean ADD Acc
01 Ape	0.152
05 Can	0.583
06 Cat	0.227
08 Driller	0.692
09 Duck	0.151
10 Eggbox	0.825
11 Glue	0.620
12 Holepuncher	0.468
Average	0.464

Reference:

《Estimating 6D Pose From Localizing Designated Surface Keypoints》

https://github.com/ecr23xx/kp6d

https://github.com/hz-ants/betapose