Day #167 - YoFlow and Dock

Reading #1 - YoFlow
YoFlow - Link1
Poster - Link2

Key Summary

• You Look Only Once (Yolo) Implemented in Tensorflow
• K-means clustering across short rolling windows to group similar objects across frames
• Predict if object is similar between frames

Yolo Basics

• Single Feedforward network
• Yolo reframe object detection as single regression problem
• Images -> Pixels -> Bounding Box -> Probabilities
• Divides image into S X S Grid
• If Object centre falls in the Grid then grid cell is responsible for detecting the object
• Predict bounding boxes and class probabilities

Yolo Implementation

• 32 layer Deep CNN
• Input Image resized into 448 x 448
• Yolo Network Output = S x S x (B*5 + C) tensor of predictions ( 7 x 7 x 30 )
• S - Numbers of rows and columns in which we divide the image
• B - Number of objects that can be predicted in given box
• C - Number of classes
• 5 - Terms account for x-axis grid offset, y-axis grid offset, width, height and confidence in each grid cell
• Dataset - Pascal VOC Dataset

K-Means Extension to Yolo

• K-means clustering across images within the short rolling window to group similar objects across frames
• Define Distance between two images I1, I2 given dimensions x,y and color channels c
• Works well when images are similar

Yoflow Weights - https://github.com/johnwlambert/YoloTensorFlow229

Reading #2 - DOCK: Detecting Objects by transferring Common-sense Knowledge
Key Summary

• Transfer Learning to transfer knowledge from source categories
• Encode similaruty, spatial, attribute and scene
• Model Objects at region level
• Leverage pretrained object detectors
• Similarity at Region Level
• Context used for Object Detection, Semantic Segmentation and Object Discovery
• Use External Knowledge to interpret common sense
• Pascal VOC Dataset

Architecture

• Base Network CNN + Common Sense
• Classification Matrix, Common Sense Matrix
• Image with Region proposals
• Common-sense cues based on similarity, attributes, spatial relations

Happy Mastering DL!!!