1. During this week, I changed the 3D convolution neural network part, and added the resnet work into it. I tried to train the model again from the first epoch, not using the pretrained pth files.

Also, I figure out the problem that the computational quantity of 3D CNN. In … Continue Reading ››

1. I moved the house to living in the GCC, and thanks for Gao Yuliang's help.
2. As for the paper I read PV-RCNN, I try to train the model based on the pretrained model. The results shows here:

Cyclist AP@0.50, 0.50, 0.50:
bbox AP:90.0048, 81.0334, 79.9002
bev AP:86.3942, 75.1885, 73.6402
3d AP:86.1937, 74.8806, … Continue Reading ››

PV-RCNN is a 3D detection paper that I am reading. This paper concentrated the 2 points into one framework: voxel CNN, and key-point refine, which is also a two-stage framework.

Last week, I continued to learn some basic deep learning network knowledge, and made a note.

I did not do some other experiments last week.

1. I used many time to learn the detailed deep learning strategy， including the classic learning structure: resnet and inception. I want to do some structure improvement in the traditional deep learning frameworks. But until now, I have no ideas when I face yolov4.
2. I used yolov5 to train the previous 3D detection … Continue Reading ››
   

Based on the work of ICISIP2021 paper, I felt the detection FPS is not so fast to maintain the real-time requirement. Actually, My 3D_Detection_YOLOv4 work could maintain 20-30 FPS on 2060 Super. The safe FPS is better above 30FPS stably. I guess my work last month could get better FPS on other machine.

This month, I put most of my time on C++ programming learning. I am learning a set of courses on the website. The course is divided by 3 parts, I have learnt all the classes except the last parts.

The powerpoint of ICISIP2021 has … Continue Reading ››

Last week I learned the Pointnet work, then this week I resume to stuty the Pointnet++. The 3D object detection framework PointRCNN used the improved pointnet++ to process the point clouds data. This is a two stage detecting work, and I am studying its program. I tried to run PointRCNN, but not … Continue Reading ››

Last week, Hong san has checked my ICISIP paper, and find out many places to revise. I have revised the paper and submitted it to ICISIP conference.

Then I started to do my later research. The previous research mainly transformed the point clouds data into 2D images. I want to … Continue Reading ››

I complete the ICISIP2021 paper last week, and git pushed to the gitlab, but still left a little part need to be modified and polished.