openpose

[nnop]

code

• http://www.cnblogs.com/caffeaoto/p/7212281.html

Code分析： testing/src/connect56LIneVec.m 意从 PAFs + heatmaps 得到连接顺序 首先是通过 func: findPeaks() 得到每张heatmaps中的亮点坐标以及对应像素点大小 score 。 然后就是计算 limbSeq 中任意一组两点的PAF值 paf_score ，然后根据paf_score进行排序，从高到低连接两点（每次连接的两点在之后的连接中都排除不用） 建立subset ，记录下从limbSeq(1) 开始两两连接点的拓展

[nnop]

data

• 姿态估计的两个数据集COCO和MPII的认识

关于关节的顺序，参考issues:

• 7
• 30 @ keras_Realtime_Multi-Person_Pose_Estimation

[nnop]

about np=56

The COCO datasets have 17 body parts, so np_in_lmdb=17.

In cpm_data_transformer from 182~184,

if(np == 56){
int COCO_to_ours_1[18] = {1,6, 7,9,11, 6,8,10, 13,15,17, 12,14,16, 3,2,5,4};
int COCO_to_ours_2[18] = {1,7, 7,9,11, 6,8,10, 13,15,17, 12,14,16, 3,2,5,4};

the 6-‘left shoulder’ and 7-‘right shoulder’ generate a new part. So there is 18 heatmaps.

In cpm_data_transformer from 1172~1173,

int mid_1[19] = {2, 9, 10, 2, 12, 13, 2, 3, 4, 3, 2, 6, 7, 6, 2, 1, 1, 15, 16};
int mid_2[19] = {9, 10, 11, 12, 13, 14, 3, 4, 5, 17, 6, 7, 8, 18, 1, 15, 16, 17, 18};

there are 19 correspondences,as well as 19*2=38 vecmaps.

Finally, 18+38=56.

Ref issues:

• 8
• 93

[nnop]

about isVisible

• 0 - occluded
• 1 - visible
• 2 - out of bounds

Ref: issues/29 @ convolutional-pose-machines-release

[nnop]

about use ResNet:

• 131 - Model with other backbone networks?

[nnop]

data augmentation

• python ref: https://github.com/anatolix/keras_Realtime_Multi-Person_Pose_Estimation

[nnop]

代码里有不少bugs, anatolix指出了一些

• 102
• 118