Mask R-CNN parse

Paper ideas

     

      Mask -RCNNFirst pictures do testing, to identify the ROI in the image, pixel correction using ROIAlign each ROI, and an ROI for each frame FCN designed using different prediction Categorize example, examples of the finally obtained image segmentation result.Examples of the difficulty is that the first division of the picture all the targets correctly detect but also to segment each example.

    *Faster R-CNN-based framework, classification and regression with the last layer, increasing the segmentation task instance [a small FCN applied to each RoI]

    *Using as a basic network ResNet101

    *RPN using 5 scales of the anchor and the 3 aspect ratios

    *The RoI Pooling Faster is replaced by R-CNN operation RoI Align

    *The final layer features using FPN (Feature Pyramid Network) feature extraction

Mask R-CNN characteristics

*Examples of segmentation

Mask -RCNN in FasterRCNN last expanded classification and regression tasks, add a task division for each RoI region. The task is a simple FCN network.

*RoIAlign operation

Because RoIPool operation, too rough, leading to a corresponding relationship between the original image and the feature level error is too large is the main problem [Fast / Faster R-CNN's], it is proposed RoIAlign operation, you can keep the precision preserves exact spatial position of [ spatial locations].

This operation, but modified a little bit, but the effect is very large, can be improved about 10% to 50% of segmentation accuracy.

*Decoupling

The segmentation and classification tasks decoupling task.

RoI classification to classify branch prediction.

FCN pixel-level multi-class classification prediction [division], including both segmentation and classification tasks.

FCN a final mask K output layer, each layer as a class, Log output is binarized to produce background and foreground segmentation Mask with 0.5 as a threshold value.

*flexibility

After the frame is very small changes, you can be human pose estimation.

The human body is a keypoint each category as a training and testing.

*time

Because the algorithm adds a very small tasks on Faster RCNN, calculate the amount of the increase is very small, which can reach the speed of 5fps.

Mask -RCNN Introduction

    

Network architecture

    

Implementation details