Happy New Year！

During this week, I thought about the next work direction. The demands for autonomous driving are multifarious，only 3D detection may not maintain the requirements. I want to expand multi-tasks for my work.

I read a multi-tasks paper "YOLOP: You Only Look Once for … Continue Reading ››

I finished all the work for AROB2022, and submitted the final paper.

I git all the files of AROB2022 to the gitlab.

AROB：I finished the paper writing of this arob2022 conference. But still need to mofidified.

Paper reading: I try to figure out some backbone network structure, like DLA-34. So I read the paper "Deep layer aggregation"

AROB:

I almost have finished the experiments of my proposed work on AROB 2022. And I got some improvement on the work.

The proposed work tend to have some improvement on pedestrian detection accuracy.

And the total detection accuracy also improved a little … Continue Reading ››

1. The improved centernet based 3D detection. The training result is shown below, on the BEV AP is not so good as the base line. I am now thinking of the reason.

car_detection AP: 97.058861 87.907043 79.321228

car_orientation AP: 93.189293 83.370262 74.603004

car_detection_ground AP: … Continue Reading ››

I finished modified the rotation yaw regression method. And the program can be succeed trained. I trained the 55 epochs.

This figure is the training loss decreasing chart.

I did the evaluation work on … Continue Reading ››

I modified the centernet 3D detection program this week. I want to change the orientation yaw regression methods. The original methods made the orientation yaw into 8 variables, which is a very complex method. Why not use the more normal and simple ways to made the yaw into 2 variables, in polar coordination … Continue Reading ››

I was arranged to do a paper-reading representation this week.

I researched the monocular 3D object frameworks, especially the keypoint-based functions: CornerNet and CenterNet.

The CenterNet 3D object detection used the CornerNet keypoint detection backborn(Hourglass network), which shows a great performance on keypoint detection. I want to make some improvements on the orientation detection, which used 8 variables in CenterNet. … Continue Reading ››

Last week，I try to make 3D object detection jobs on Monocular camera, not using lidar datas.

I read some monocular papers and tried some opensource programs. I used the my own experiment data to do some test experiments on centernet, which is also a monocular camera based open program. The … Continue Reading ››