Research/Blog
#AIByte by CellStrat – Object Detection with Mask R-CNN
- November 13, 2019
- Posted by: vsinghal
- Category: Computer Vision Deep Learning
Region CNNs (RCNNs) and their variations such as Fast RCNN, Faster RCNN and Mask RCNNs are state-of-the-art algorithms for Object Detection in images.
RCNNs detect 2000 region proposals using Selective Search, compute CNN features on each Region, and then classify each region as certain objects or not. The final layer also does bounding box regression to create bounding boxes around the objects.
Fast RCNN adds a preliminary CNN layer to create features to be shared with Region proposal method. The output of Region proposals is fed to a new RoIPool layer which pools the RoIs, and then finally a softmax classifier specifies the objects in RoIs and bounding box regressor predicts the BB coordinates.
Faster RCNN adds a Region Proposal Network (RPN) to a Fast RCNN model (instead of Selective Search) to propose regions more accurately. RPN takes image feature maps as an input and generates a set of rectangular object proposals, each with an objectness score, as output. It uses Anchor boxes that serve as reference boxes at multiple scales and aspect ratios, to be evaluated, in order to detect regions containing objects.
Mask RCNN adds a parallel branch to a Faster RCNN to additionally predict object masks for each Region of Interest (RoI). It uses a RoIAlign layer maintaining pixel-level spatial intelligence during RoIPool, without which the segmentation mask would not be possible.
Fast RCNNs, Faster RCNNs and Mask RCNNs use a multi-task loss minimization approach, where different kinds of losses are summed up in weighted manner and the cumulative loss is minimized.
Mask RCNNs represent state-of-the-art in object detection, instance segmentation and also great for pose estimation applications.
CellStrat AI Lab Workshop on “Object Detection with Mask RCNNs” on Saturday 16th Nov 2019 :-
I invite you to attend our Hands-On Workshop on Mask RCNNs this Saturday (16th Nov 2019) in BLR and Gurgaon, where we will present extensive training on various RCNN variations and also run a hands-on model on a image dataset. Register below :-
Bengaluru AI Lab :-
Register : https://www.meetup.com/Disrupt-4-0/events/vcqljryzpbvb/
Gurugram AI Lab :-
Register : https://www.meetup.com/Disrupt-4-0/events/265659070/
Join us for world-class AI disruption ! See you soon..
Best Regards,
Vivek Singhal
Co-Founder & Chief Data Scientist, CellStrat
+91-9742800566