Binocular 3D maize key point structure

1.The Backbone is transformed into a mobilenet structure, and the left and right eye images are used as inputs to extract features and generate cost-volume to obtain depth.

1.Based on the backbone of YOLOV7-POSE, propose the YOLOV7-3DPOSE to directly output the 3D keypoint.And i will use the framework to detect the keypoint of the corn.

2.The system of the yolov7-pose is complex, i am learning it.

1.make research on the code for yolo-v5 pose

2.try to change the code of yolostereo3d to make 3D keypoints detection

3.In the study of binocular images, parallax estimation is performed only after the BBOX is detected，and then the darta source will be no different from RGB-D.

1.Revise my paper

2.Make research on the 6d object detection based on the template. (we need constructing the 3D model to complete the project)

related paper:POSE CNN,SSD 6D,DEEP-6D POSE

comparison： get the corresponding projection points from the regression.And … Continue Reading ››

1.read the paper about 6D pose emstimation(A Review of 6D Object Pose Estimation).

try to use the 6D POSE emtimation to find the corns in the 3D coordinate system.

trying to find a method based on key-point datection to make 6D Pose Estimation

1. Finish the model Stereo-FCOS-3D(2023/2-3)
2. use the dataset to test the model(2023/3)
3. Finish the paper Stereo-FCOS-3D(2023/4-5)

I'm waiting for the corn to ear, and then I'm about to start my new project.

Uniformer（https://openreview.net/forum?id=nBU_u6DLvoK）:

Introduction:

Combining the ability of aggregation feature context of 3D convolution with the self-attention mechanism of visual transformer, to improve the ability of remote dependence of the model. Uniformer combines the advantages of both and achieves good results in many visual tasks.

1.read 80% of paper(" Study on Deep Learning Based").

2.use depth prior to guidance the model is one important focus in paper(https://link.springer.com/article/10.1007/s00371-022-02607-x) and Study on Deep Learning Based. The next work is to learn the code of the https://github.com/xjtuwh/iDepNet/tree/master.

3.make research on … Continue Reading ››

1.Learn the context fusion model https://arxiv.org/pdf/1912.06319.pdf.

2.In fact, I think, the context mechanism doesnt has much difference with Multiscale fusion with ghost moudle.

3.Based on IDA-3D,i will use the context fusion mechanism to improve the whole module. Regarding the detailed structure of the … Continue Reading ››

1.Successfully configure the environment of DSGN++.

2.The method of DSGN is to transform the information of binocular image into three-dimensional space, and keep the color and texture information. (In contrast, the pseudo-radar method does not retain color and texture information, but directly generates point clouds)

3.I prefer … Continue Reading ››