0. Problem definition, Main Challenges

<aside> 🧑‍🏫 WSOD aims to classify and locate object instances using only image-level labels in the training phase.

</aside>

Main Challenges

1. Discriminative Region Problem
   • detectors tend to focus on the most discriminative parts of the object
2. Multiple Instance Problem
   • Detectors are difficult to accurately recognize multiple instances when there may exist everal objects with the same category in an image
3. Speed Problem

1. Basic WSOD Framework

MIL-based Networks

• Multiple instance prediction → MIL(multiple instance learning) problem

• Based on “Weakly Supervised Deep Detection Network” which consists of the following

1. Proposal Generator
   • Selective Search
   • Edge boxes
   • Sliding Window
2. Backbone
3. Detection Head

CAM-based Network

• Single Instance prediction → Object Localization Problem

1. Backbone