目标检测串烧

图像基本问题

COCO数据集与评价指标

评价指标

例子

model_name='fasterrcnn_resnet50_fpn' params=41755286 duration='0:09:07
IoU metric: bbox
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.36934
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.58546
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.39625
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.21201
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.40316
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.48154
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.30748
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.48482
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.50857
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.31752
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.54431
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.64890

Naive Approach

检测框损失及度量

Non-maximum Suppression (NMS)

方法

Region CNN (RCNN)

  1. 提候选框 ~2k
  2. 缩放成统一尺寸
  3. CNN 抽特征
  4. SVM分类器, 回归器修正候选框位置

TODO 传统图像分割算法 / Region Proposal Algorithms

Spatial Pyramid Pooling (SPP)

Fast RCNN

  1. 全图特征过CNN提取特征图 (feature map), 避免候选框特征提取的重复计算
  2. ROI pooling 层将特征图 (SPP net 简化版)
  3. FC分类 + FC回归 一起训练
  4. 相对SPP优化, 可以同时训练CNN和FC层

Faster RCNN and Region Proposal Network (RPN)

Mask RCNN 增加了目标分割的FC分类输出训练

Feature Pyramid Networks (FPN)

YOLO

Single Shot Multiscale Detection (SSD)

RetinaNet & Focal Loss

YOLOv3

DETR / End-to-End Object Detection with Transformers

二分图匹配问题

Deformable DETR

2022.7 YOLOv7

再次刷榜, 需要关注

模型评价对比

业务场景

HOME