2019 第五届“四维图新”杯创新大赛 自动驾驶视觉综合感知算法赛 detection baseline

November 6, 2019 · View on GitHub

testA object detection map: 0.22

testA semantic segmentation miou: 0.46

author: zhengye

for the semantic segmentation part, please refer to my teammate's git repo: https://github.com/SHERLOCKLS/datafountain_siweituxin_autodriver_seg

环境配置

请按照mmdetection说明进行安装配置

训练

  • 数据准备

    将训练数据所有训练图片放置于 data/siweituxin/train_image

    两批次的数据label分别位于'data/dataset1/train.txt' 和 'data/DF_1018/train.txt'

    合并label: python tools/convert_datasets/merage_txt_label.py

    txt转coco json: python tools/convert_datasets/trans_txt2json.py

  • 模型训练

    请依据mmdetection说明依据自身显卡情况线性调整lr,这里以4卡为例

    CUDA_VISIBLE_DEVICES=0,1,2,3 ./tools/dist_train.sh config/siweituxin/faster_rcnn_r50_fpn.py 4

预测

  • 数据准备

    将测试数据所有训练图片放置于 data/siweituxin/test_images

  • inference

    CUDA_VISIBLE_DEVICES=0 python tools/infer_siweituxin.py config/siweituxin/faster_rcnn_r50_fpn.py /work_dirs/faster_rcnn_r50_fpn/latest.pth --out det.txt

Contact

This repo is currently maintained by Ye Zheng (@zhengye1995).