April 2021
tl;dr: Automatically adapt gaussian kernel size for keypoint detection to accommodate scale variation and labeling ambiguity.
Overall impression
The paper addresses one key issue in bottom up key point detection.
Key ideas
- SAHR (scale adaptive heatmap regression) predict an additional scale maps s to modify the standard deviation to be $\sigma_0 s$. This improves the baseline of HPE (human pose estimation) which covers all keypoints by Gaussian kernels with the same stdev.
- Loss: L2 heatmap regression loss + regularization loss
- The size map can also be interpreted as uncertainty.
- Regularizer loss has to be added to stablize training. This is more like the aleatoric uncertainty estimation.
- WAHR (weight adaptive heatmap regression): more like focal loss used by CornerNet and CenterNet. It automatically downweighs the loss of easier samples.
- Suppose P is prediction, H is GT heatmap
-
original l2 loss $ |
|
P-H |
|
_2^2$ |
-
new loss $W |
|
P-H |
|
_2^2$, and $W = H^\gamma |
|
1-P |
|
+ (1-H^\gamma) |
|
P |
|
$. |
-
Intuitively, for points larger than a certain thresh, the weight is more like $ |
|
1-P |
|
$ and penalizes more if the prediction is small. |
Technical details
- Bottom-up vs top-down approaches for human keypoint detection.
- Bottom up approaches could be faster as the running speed is not limited by the number of instances, and is not limited by the performance of the object detector. See PifPaf. Keypoints are grouped by Associative Embedding.
- Top down approaches are two-stage methods. Images are cropped out, resized before feeding into a single-human keypoint detector. The inference time scales with number of people in the image. Currently top-down approaches still have better performance, except in crowded scenes.
- Each keypoint has a scale map s, to pair with the heatmap.
- Ablation study
- Naive implementation is to estimate size of heatmap from bbox size. This is actually worse than the baseline performance with a uniform stdev.
Notes