Skip to content

阿里巴巴高德地图AMAP-TECH算法大赛 - 基于车载视频图像的动态路况分析https://tianchi.aliyun.com/competition/entrance/531809/introduction

Notifications You must be signed in to change notification settings

ALLinLLM/gaode_congest

Repository files navigation

天池高德比赛

提交记录

模型 分数 备注
lightgbm 0.64 根据视频帧的时间,人工设计一系列特征
cos_sim 0.5155 VGG19(conv2d-19) + argmin(1-cos(a,b))
vgg19feature_extract - 过拟合

data

参考数据目录datasets/README.md

startup

  1. git clone git@github.com:vegetable09/gaode_congest.git
  2. cd gaode_congest
  3. 按照datasets/README.md下载好数据和json到datasets/
  4. cd codes/data
  5. sh get_Xy.sh
  6. 下载https://download.pytorch.org/models/vgg19-dcbb9e9d.pth~/.cache/torch/checkpoints/vgg19-dcbb9e9d.pth
  7. sh run_model_cos.sh
  8. python embedding_distance_test.py
  9. 结果保存在datasets/amap_traffic_annotations_test_result.json

codes

所有方法在codes/models目录, 目前有

  • lightgbm: 根据视频帧的时间,人工设计一系列特征
  • cos_sim: 使用VGG19的卷积层获得每个分类的embedding中心, 对于test数据, 计算余弦距离, 取最近的
  • vgg19feature_extract: 使用完整的VGG19, 接上softmax层, 过拟合

About

阿里巴巴高德地图AMAP-TECH算法大赛 - 基于车载视频图像的动态路况分析https://tianchi.aliyun.com/competition/entrance/531809/introduction

Resources

Stars

Watchers

Forks

Packages

No packages published