aiacademy: 深度學習 CNN YOLO(You only Look Once)

YouTuber : Siraj Raval

• YOLO 從darknet 寫的，網路上大神改寫最好的叫 darkflow

Yolo (You only Look Once)

Yolo2

Yolo Structure

We frame object detection as a regression problem to spatially separated bounding boxes and associated class probabilities

YOLO: Limitation & Generalization of results

Limitation:

1. 一個 grid 只預測 2 個bbox, 而且只會有一類　=> 靠很近的 Object detect 效果不好

2. 同一類 Object 新的長寬比例效果不好

3. 小　Object 的 localization error

YOLO9000: Better, Faster, Stronger

	Description
Batch Normalization	Add BN after every conv, delete dropout
High Resolution Classifier	Finetune on Imagenet as a classifier with 448 x 448
Convolutional With Anchor Boxes	YOLO: S x S x (B x 5 + C), YOLOv2: S x S (B x (5 +C))
Dimension Clusters	Kmeans for Anchor Boxes
Direct location prediction	Direct predict, normalize to 0 ~ 1
Fine-Grained Features	Add pass-through layer
Multi-Scale Training	Random change input shape after a number of epochs

YOLO2 performance and mAP

• AP: average performance

YOLOv2 implement

• GitHub

• Pascal VOC

  • 20 classe
    • person
    • bird, cat, cow, dog, horse, sheep
    • aeroplane, bicycle, boat, bus, car, motorbike, train,
    • bottle, chair, cining table, potted plant, sofa, tv/monitor

  • https://blog.csdn.net/weixin_35653315/article/details/71028523

  • xml 紀錄圖片內容的訊息

    

• Microsoft COCO

  • 80 classes
  • https://blog.csdn.net/u012905422/article/details/52372755

YOLOv3

Class Prediction