【CrowdHuman】《CrowdHuman：A Benchmark for Detecting Human in a Crowd》

arXiv-2018

---

---

1 Background and Motivation

现有人体检测公开数据集样本不够密集，遮挡也不够

Our goal is to push the boundary of human detection by specifically targeting the challenging crowd scenarios.

于是作者开源了一个密集场景的人体检测数据集

2 Related Work

• Human detection datasets
  exhaustively annotating crowd regions is incredibly difficult and time consuming.
• Human detection frameworks

3 Advantages / Contributions

开源了一个 larger-scale with much higher crowdness 的行人数据集——CrowdHuman，兼具 full body bounding box, the visible bounding box, and the head bounding box 标签，实验发现是一个强有力的预训练数据集

4 CrowdHuman Dataset

4.1 Data Collection

Google image search engine with ∼ 150 keywords for query.

搜索的关键字涵盖 40 different cities，various activities，numerous viewpoints，比如 Pedestrians on the Fifth Avenue

a keyword is limited to 500 to make the distribution of images balanced.

爬下来 ~2.5W 张，整理

15000, 4370 and 5000 images for training, validation, and testing respectively.

4.2 Image Annotation

先标 full bounding box

把 full bbox 裁剪出来，再标 visible bounding box 和 head bounding box

4.3 Dataset Statistics

Dataset Size / Diversity

Density

Occlusion

visible ratio 越小表示遮挡越严重，极限遮挡的话 CityPersons 还是会比 CrowdHuman 多一些

除了上面的二人遮挡，作者还统计了三人遮挡率

5 Experiments

检测器 FPN and RetinaNet

5.1 Datasets and Metrics

• Caltech dataset

• COCOPersons，64115 images from the trainval minus minival for training, and the other 2639 images from minival for validation.

• CityPersons

• Brainwash

• Recall

• AP

• mMR，which is the average log miss rate over false positives per-image ranging in $[10^{-2},10^0]$，越小越好

5.2 Detection results on CrowdHuman

先看看 visible bounding box 的检测结果

看看可见部分的检测示例

再看看 full bounding box 的检测结果

再看看人头检测

5.3 Cross-dataset Evaluation

看看其泛化性能

COCOPersons

可以看到用 CrowdHuman 预训练过后，再在 COCOPersons 上微调效果有提升

Caltech

CityPersons

Brainwash

6 Conclusion（own）

https://github.com/sshao0516/CrowdHuman