Skip to content

Latest commit

 

History

History
60 lines (51 loc) · 2.13 KB

README.md

File metadata and controls

60 lines (51 loc) · 2.13 KB
Raw

Overview of My Project

Alt text

Conduct of the Code

Structure

  • All my own codes are in mywork.ipynb
  • The segment_anything and scripts come from the project Segment Anything and if you want to explore them more can visit https://github.com/facebookresearch/segment-anything
  • sam_vit_h is a checkpoint
  • Images include all test picture
  • outputpicture include all rebuild pictures from all the masks extracted from the original pictures
  • 9mask includes all masks from the ninth picture and as I mentioned in the report it is for proving the great segmenting ability of the SAM
  • More work includes some code of the "More Advanced Works" in the report

How to Run

Environment

  • opencv, pytorch, numpy and so on
  • Other import environment can refer the Segment website and I also include their README in Segment Readme
  • Notice! This work needs GUP resource otherwise maybe OOM

The main part of the code is in the last cell of the mywork.ipynb and you can just run it to see the output , of courese, allowing the upper parts having been made

some parameters

The SAM

#掩码生成器参数修改
mask_generator_2 = SamAutomaticMaskGenerator(
    model=sam,
    points_per_side=20,#控制采样点的间隔，值越小，采样点越密集
    pred_iou_thresh=0.86,#mask的iou阈值
    stability_score_thresh=0.96,#mask的稳定性阈值
    crop_n_layers=1,
    crop_n_points_downscale_factor=2,
    min_mask_region_area=50,  #最小mask面积，会使用opencv滤除掉小面积的区域
)

The Filiter

    #统计汉明距离小于 5 的图像对
    for i in range(n):
        #j从i+1开始，因为相似度矩阵是对称的
        for j in range(i+1,n):
            if hash_matrix[i, j] < 8:#调节这个阈值
                images.pop(j)
                image_areas.pop(j)
                n -= 1

How many catagories you want

clusters = k_means_clustering(areas1,10)#所求类别
print(clusters)

Other parts have a lot of Chinese notices and you can explore them

Thanks

  • Professor Zhao, tutor Chen and the meaningful knowledge in DIP class
  • Meta and their fantastic work SAM