Mask R-CNN for Human Pose Estimation

Note

• This code has not been cleaned, and still may be difficult to read.
• The evaluation part has not been tested, it may come to crash in test mode.

Configuration

• Please refer to the original version matterport/Mask_RCNN to set up the environments.
• Modify the keras code in lines 2552-2353 of evaluate_generator() in the file /engine/training.py: averages.append(np.average([out[i] for out in all_outs], weights=batch_sizes, axis=0)). Or you can update the loss code as recently updated in matterport/Mask_RCNN

Pitfalls

• When using the horizontal flip augmentation, be aware of the "mirror person" problem. Maybe we should not use this kind of augmentation.

Related Works

• RodrigoGantier/Mask_R_CNN_Keypoints
• Superlee506/Mask_RCNN_Humanpose It has a more efficient space encoding scheme when calculating the kp mask loss.