I continued training the next epochs of the 3D detection framework created in last week.

I made the 3D detection framework to be able to detect videos, and this 3D detection work is proved to have the high detect accurate and also can maintain the real-time detecting demand.

As for the research mind for last week, I changed the previous framework from a 2D detection to 3D detection. Actually, the previous only did a 2D regression task, but and build a 3D detection bounding boxes. I changed the framework to a real 3D regression task(x,y,z,h,w,l,yaw), but the detection precision becomes lower … Continue Reading ››

Last week, I used a optimized yolov4 network to improved the detection accuracy on 3D detection.

While, the 3D object detection framework only regressed 5 parameters( x, y, w, h, yaw). Since we want to detect a 3D bounding box, I decided to use yolo to directly regress all the … Continue Reading ››

I tried many networks based on yolov4 for 3D detecting.

complex_yolo_tiny 180 epoches 0.69975

complex_yolo_tiny 250 epoches 0.74941

complex_yolo 30 epoches 0.637739

complex_yolo 40 epoches 0.692167

complex_yolo 50 epoches 0.7347

improved_yolov4 … Continue Reading ››

I will do a presentation in this week.

I researched the 3D detection fuctions of yolov4, which I mentioned before. This project used Bird eye view of point cloud map, and build 2D bounding boxes both on the image and point cloud maps. Finally, mapping the 3D box result on the images. This function actually trained only 4 variables for a … Continue Reading ››

I found an idea that using yolo to detect real-time 3D bounding boxes, and I also found the existing yolo_based 3D object detection algorithms:

Complex-YOLOv2, Complex-YOLOv3 and Complex--YOLOv4

I trained the Complex-YOLOv4 without using pretrained .pth file, and trained for 50 … Continue Reading ››

I mainly submitted the application form of DC2, and also modified the research plan form.

I mainly modified the two points:

1. why I come up with this research topic.

2. And the contribution of my research topic to society.

I researched the yolov1-yolov3, and plan to use yolo to build a 3D object detection with kitti dataset. And I have found a project using yolov4 to build a 3D object detection system.

I followed the Waymo dataset benchmark, and overviewed the outstanding frameworks on it. I recently followed a soon graduating doctoral student, and his works inspired me a lot.