June 2020
tl;dr: Monocular estimation of body orientation.
The paper annotates body orientation in COCO dataset.
TUM dataset has 8 bins, and got later extended to continuous labels by averaging annotations from 5 labelers.
- 72 bins x 5 deg interval for annotation. This is still within the human cognition limit.
- The circular gaussian loss is used to blur the one-hot classes and then regresses a heatmap. Interestingly the paper uses L2 loss directly instead of a BCE loss.
$p \propto e^{-\frac{1}{2\sigma^2}(\min(|i-gt|, 72-|i-gt|))}$ - The loss is approximation von Mises distribution (see mono3D++ which actually used this idea to regress angles in 360 bins).
- The paper also formalized the definition of body orientation by TxS, perpendicular to both torso direction and shoulder direction. This is needed to incorporate body orientation as a weak supervision to 3d body estimation.
- Labeling tool design. It is a brilliant idea to show case example images with body orientation. (This could be done for car orientation estimation as well).
- The version with appendix can be found on amazon.
- Classification over regression: Sometimes it is useful to convert regression to a multi-class classification with ordered bins. Instead of directly predicting a one-hot label, it blurs the one hot label by allowing leakage into its neighboring bins. This is similar to the idea of label smoothing to stabilize training. I feel that this is actually a particularly useful technique for multi-bin classification, when the bin numbers are very large, and when bins are ordered.
- Designing Deep Convolutional Neural Networks for Continuous Object Orientation Estimation proposed three different methods to regress a continuous angle.