Skip to content

Latest commit

 

History

History
45 lines (32 loc) · 2.15 KB

model_README.md

File metadata and controls

45 lines (32 loc) · 2.15 KB
Raw

Didi challenge

How to run train.py

Store kitti dataset in data_dir (line 112 file train.py). This folder is orgalnized as follows:

data_dir

  • 0001
    • gt_boxes3d
      • gt_boxes3d_0.npy
      • gt_boxes3d_1.npy
      • ...
    • lidar
      • lidar_0.npy
      • lidar_1.npy
      • ...
  • 0002
    • ...

The function list_of_data() (lines 13-26) creates a two lists. One contains names of lidar files and other contains names of ground true boxes.

The function train_batch_generator() creates a generator for training.

The trained weight will be stored as ./saved_model/model.h5.

The model

Please refer to the paper Vehicle Detection from 3D Lidar Using Fully Convolutional Network for the detail.

The ideal of the model as follows:

Input : tensor of shape (64,256,2), which is the 84 degree panoramic view of lidar, contains the depth of height information of lidar point.

Output: two tensors:

  • A tensor of shape (64,256,1) where the number at the position [x,y,0] presents the at which probability the the pixel [x,y] belongs the the obstacle. This part works as a segmentation problem (out1 line 110, file fully_conv_model_for_lidar.py)
  • A tensor of shape (64,256,7) where the first 6 value vector [x,y,:] presents the coordinate of two opposite corners of the bounding box containing the point [x,y], the last value encodes how narrow the bottom rectangular of the box is. I chose the corners number 0 and 6 of the box, the other six corners will be inferred from those values. This part works as a multi-regression problem (out2 line 118, file fully_conv_model_for_lidar.py)

Loss function: function my_loss() line 88-107 file train.py. The total loss includes two losses: log loss for segmentation part (seg_loss) and mean squared error (MSE) for regression part.

Prediction and clustering the output boxes

The function predict_boxes() in file util_func.py.

Input: the trained model and lidar points (of shape N*C, C>=3)

Output:

  • all_boxes: all of boxes with segmentation probability >= 0.5
  • cluster_boxes: list of boxes after clustering