D-DL4CV-Lec15-Object_Detection

1. Object Detection Introduction

1.1 Task Definition

Input: Single RGB image

Output: A set of detected objects with:

1. Category label (from fixed, known set of categories)
2. Bounding boxes (x, y, width, height)

1.2 Challenges

• Multiple Output: The number of objects varies
• Multiple types of output: The task require model to not only detects “what” is the object but also “where” is the object
• Large Images: Images contains multiple objects tend to have high resolutions

---

2. Detecting a Single Object

D-DL4CV-Lec15a-Single_Object_Detection

---

3. Detecting Multiple Objects

3.1 Sliding Windows (Bad Approach)

D-DL4CV-Lec15b-Detecting_Multiple_Objects_with_Sliding_Windows

3.2 Region Proposals

3.2.1 Method

We create some algorithms that can find a small set of boxes that are likely to cover objects in the image, then we process classification on these boxes

3.2.1 R-CNN

D-DL4CV-Lec15c-R-CNN

• Intersection over Union (IoU)
• Non-Max Suppression (NMS)
• Mean Average Precision (mAP)

3.2.2 Fast R-CNN

D-DL4CV-Lec15d-Fast_R-CNN

• Rol Pooling
• Rol Align

3.2.3 Faster R-CNN

D-DL4CV-Lec15e-Faster_R-CNN

• Region Proposal Network (RPN)

3.3 Single-Stage Object Detection

D-DL4CV-Lec15f-Single-Stage_Object_Detection