Idea:
Use the existing yolostero3D framework to optimize and improve through novel cost_volume and DSGN++（https://arxiv.org/abs/2204.03039） methods.And make own dataset for training and testing in the future.
Other idea:
Adopt newer yolo ideas (yolo7 yolox, etc.)
Novel cost volume build - reference GA-NET（https://arxiv.org/pdf/1904.06587v1.pdf） … Continue Reading ››
1.Practice:
Run semantic segmentation based on pytorch, and make a simple test data set.
Multi-spectral image processing and network model building.
2.Learn anchor-free ideas and codes.
1.Update Yolostereo3D model based on attention mechanism
2.read paper
Adaptive unimodal cost volume filtering for deep stereo matching,https://ojs.aaai.org/index.php/AAAI/article/view/6991/6845
3.Disassembly code of yolostereo3d
1.After making research on the code of yolostereo3D,i found Pyramid is the basic of yolostereo3D,i am learning it.
2.Learn ocr, my intership need it.
1.FADNEThttps://github.com/HKBU-HPML/FADNet
make research on the code
2.stereo match is a subframe of FADNET, i am trying to learn the detail of stereo matchhttps://www.cv-foundation.org/openaccess/content_cvpr_2016/html/Luo_Efficient_Deep_Learning_CVPR_2016_paper.html
3. maybe the improvement of stereo match can improve the 3D object detect
1.setup PV-RCNN 2.read paper(https://arxiv.org/abs/2103.09422) YOLOStereo3D
1.Based on NAMhttps://arxiv.org/abs/2111.12419 and Shuffle Attention(https://arxiv.org/pdf/2102.00240.pdf) ，I propose my own attention mechanism CWSG(channel weight spatial group).
Its result is 93.64 almost like SA attention.
2.Next aim is to use multi-model combination to get the final prediction result.
3.I am writing my … Continue Reading ››
1.8 classes result:
SLIR with SAM:94.2
SLIR:93.51
INV2:92.9
Resnet50:92.9
2.Prepare to write a graduation thesis based on the above data.
1.Add two new categories through crawler, and start doing experiments.
2. Template and outline of graduation thesis have been prepared and Prepare to start writing graduation thesis.
3.Finish the Arob Paper.
1.Learning code of point cloud filtering and downsampling
2.Completed the first draft PPT for doctoral admission.