3D Bounding Box Estimation Using Deep Learning and Geometry

Introduction

This repo is PyTorch implementation for this paper. In this paper, they collect KITTI 2D Object Dataset and introduce a flow to estimate object pose and dimension. If you are looking for TensorFlow implementation, here is a great repo.

Dependency

numpy
opencv
yaml
PyTorch
torchvision
CUDA

Usage

Before using this code, you need download data from KITTI and unzip it. After that, you need to add the kitti path of dataset to config.yaml.

kitti_path: somewhere # Root of kitti, where contrain trainning/ and testing/

Also, you can set up parameters for training and weight of loss as describded in paper.

epochs: 8 # How many epoch for training?
bins: 2  # How many bins you want to split?
w: 0.8
alpha: 0.8
batches: 8

After setting up, just type it for training

python Train.py

It will store model in ./models. For simple evaluation, type

python Eval.py

This will calculate average orientation and dimension error (in degree and meters).

Reference