Doodies / Person-Extraction

Real time person extraction with utmost accuracy

Objective:

Use computer vision techniques to identify and remove background from a video stream in real time.

How we tackled it:

We thoroughly researched the works in this field and came up with many different solutions that can solve the given challenge. Following are the approaches we tried:

1. YOLACT: Real time instance segmentation link
2. SIAM-MASK: Fast Online Object Tracking and Segmentation link
3. BodyPix: Person Segmentation in the Browser link
4. Mask-RCNN: Object Detection and Segmentation link
5. DeepLabs: DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs link
6. Google AI: Mobile Real-time Video Segmentation link
7. GrabCut Algorithm: Interactive Foreground Extraction link

Best Approachs:

YOLACT: Real time instance segmentation

1. Fully-convolutional model for realtime instance segmentation that achieves 29.8 mAP on MS COCO at 33 fps evaluated on a single Titan Xp, which is significantly faster than any previous competitive approach.
2. Accomplish this by breaking instance segmentation into two parallel subtasks: (1) generating a set of prototype masks and (2) predicting per-instance mask coeffi-cients.
3. Then we produce instance masks by linearly combining the prototypes with the mask coefficients. We find that because this process doesn’t depend on repooling, this approach produces very high-quality masks and exhibits temporal stability for free.
4. Also propose Fast NMS, a drop-in 12 ms faster replacement for standard NMS that only has a marginal performance penalty
5. The processing speed is very fast for real time data extraction. On Titan Xp GPU, the results are: Image Size: 550 Backbone model: Resnet50-FPN FPS: 42.5

SiaMask: Fast Online Object Tracking and Segmentation

1. perform both visual object tracking and semi-supervised video object segmentation, in real-time, with a single simple approach.
2. SiamMask, improves the offline training procedure of popular fully-convolutional Siamese approaches for object tracking by augmenting their loss with a binary segmentation task.
3. SiamMask solely relies on a single bounding box initialisation and operates online, producing class-agnostic object segmentation masks and rotated bounding boxes at 55 frames per second.
4. Speed are tested on a NVIDIA RTX 2080. FPS: 56

Other Approaches:

BodyPix: Person Segmentation in the Browser

This approach was requires GPU machine and the output was very slow fps.

Mask-RCNN: Object Detection and Segmentation

This approach had decent accuracy but the processing time was very high, approx 14 seconds per frame, so it was not feasible for real time.

DeepLabs: DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs

This approach doesn't give high accuracy and the processing was also very slow.

Google AI: Mobile Real-time Video Segmentation

The accuracy was not good in this approach.

GrabCut Algorithm: Interactive Foreground Extraction

This approach required the bounding box to be detected for each frame and then process for foreground extraction, hence this was having low fps.

Results:

Input

Output

Input

Output