hustzxd / task1-pedestriandetection

视频处理分析的task

---

---

Pedestrian detection

Video processing analysis assignment.

SSD for pedestrian detection in RAP dataset.

Train

• First download the fc-reduced VGG-16 PyTorch base network weights and trained weights at BaiduYun

• By default, we assume you have downloaded the file in the ssd.pytorch/weights dir:

cd ssd.pytorch
# train SSD300 on MS COCO person dataset
./train_coco_person.sh
# train SSD300 on Pascal VOC 2007+2012 dataset
./train_voc_person.sh
# train SSD300 on both COCO and VOC dataset
./train_voc_coco_perosn.sh

Evaluation

cd ssd.pytorch
# eval on Pascal VOC 2007
./eval_voc_person.sh weights/<you.pth>
# eval on RAP test dataset
./test_rap.sh weights/<you.pth>

Performace

We just evaluation AP of person in Pascal VOC 2007.

Training Data	Original	Only person(this project)
07+12	76.2 %	77.8%
07+12+COCO	81.4%	82.98%
07+12+COCO+RAP	—	82.49%

Reference

• Wei Liu, et al. “SSD: Single Shot MultiBox Detector.” ECCV2016.
• Thanks to amdegroot.