KAPAO is an efficient single-stage human pose estimation model that detects keypoints and poses as objects and fuses the detections to predict human poses.
I have my custom dataset so I thought calculating data specific anchors is a better idea. (using kmean_anchors).
These are the 8 values I got based on my data:
32,32, 108,195, 196,229, 159,348, 260,328, 359,317, 281,518, 509,424
I have few questions:
1- The kmeans_anchor algorithm calculates these anchors without actually doing the augmentations. Not sure if that is a good idea. What do u think ? Its currently calculated based on distribution of (width,height) of pure training data (without augmentations).
2- How should I assign these anchors to different grid ? I guess particular size of anchor boxes are more suitable to specific grid. Latest yolo5s.yaml uses three set of grids instead of 4. So I decided to use that and this is how it looks like.
I added an extra 20,20 anchor to make the total of 9 anchors. Do you think this is optimal ? I am still training the model.
3- If I only want to use 6 anchors:
32,32, 122,195, 165,314, 281,311, 281,473, 483,403
Which grid size should I skip and what to remove from head section of yolo.yaml file ?
4- Can I use a pre-trained model which was trained with different anchor boxes ? OR should I train from scratch whenever I come up with new anchors?
I have my custom dataset so I thought calculating data specific anchors is a better idea. (using kmean_anchors).
These are the 8 values I got based on my data: 32,32, 108,195, 196,229, 159,348, 260,328, 359,317, 281,518, 509,424
I have few questions: 1- The kmeans_anchor algorithm calculates these anchors without actually doing the augmentations. Not sure if that is a good idea. What do u think ? Its currently calculated based on distribution of (width,height) of pure training data (without augmentations). 2- How should I assign these anchors to different grid ? I guess particular size of anchor boxes are more suitable to specific grid. Latest yolo5s.yaml uses three set of grids instead of 4. So I decided to use that and this is how it looks like.
I added an extra 20,20 anchor to make the total of 9 anchors. Do you think this is optimal ? I am still training the model.
3- If I only want to use 6 anchors: 32,32, 122,195, 165,314, 281,311, 281,473, 483,403 Which grid size should I skip and what to remove from
headsection of yolo.yaml file ?4- Can I use a pre-trained model which was trained with different anchor boxes ? OR should I train from scratch whenever I come up with new anchors?