Implementation of gupnet on paddle3d

Special changes

|-paddle3d-gupnet(ROOT)
  |-configs
    |-gupnet_dla34_kitti.py      # 配置文件
  |-docs
    |-gupnet
      |-README.md	         # 复现说明文档
  |-paddle3d
    |-datasets
      |-kitti
        |-kitti_gupnet.py        # 数据集加载
        |-kitti_gupnet_utils.py  # 数据集加载辅助函数
        |-kitti_utils.py         # gupnet相关数据处理函数
    |-models
      |-detection
        |-gupnet
          |-gupnet.py            # 模型结构搭建
          |-gupnet_dla.py        # 特征提取网络结构
          |-gupnet_helper.py	 # 辅助函数
          |-gupnet_loss.py       # 损失
          |-gupnet_predicter.py  # 预测
          |-gupnet_processor.py  # 后处理
        |-optimizers
          |-lr_schedulers.py     # 定制WarmUp+MultiStep学习率

Prepared

1. Modify the dataset root directory (dataset_root) in the config file (configs\gupnet\gupnet_dla34_kitti.yml)

2. Download the pre training weights of dla34 to the checkpoint folder

link：https://pan.baidu.com/s/1NTqgwaBAq2jP2YJ_X8cmzw?pwd=swxl

pw：swxl

Install

Installation Tutorial

Train

on one gpu:

$ python tools/train.py --config configs/gupnet/gupnet_dla34_kitti.yml --batch_size 8 

Val

$ python tools/evaluate.py --config configs/gupnet/gupnet_dla34_kitti.yml --batch_size 8 --model output/epoch_140/model.pdparams

Modify the model parameters to the weights you have trained

Best result:

Best car performance on the validation set:

Code	Venue	Easy	Moderate*	Hard
Official	ICCV'21	23.19	16.23	13.57
repo	ICCV'21	21.44	15.44	12.84