README.md

DeepVO

This repository is an unofficial PyTorch implementation of DeepVO. Here is the link to the original ICRA 2017 paper. The codes are mainly based on another unofficial implementation of DeepVO but with some modifications and extentions.

Key differences:

• The generation of ground truth relative poses is modified. We first compute the 4x4 transformation matrix between adjacent frames and then convert it into rotations and translations. For example, we have two 4x4 transformation matrices M1 and M2 that represent the global poses (relative to the starting point) at time 1 and 2. Then, the relative pose transformation matrix M' between time 1 and 2 can be calculated by M' = inv(M1) * (M2). Then, we can extract R' and t' from M'. t' is the relative translation and the relative rotations can be extracted from R' using function euler_from_matrix in helper.py.
• The inference is different. Here we directly input the whole video to the LSTM without breaking them into independent segments. The way we accumulate the relative poses is also different.
• Horizontal flipping is added as an augmentation method

Dataset

Please download the KITTI Odometry dataset (color, 65GB) following this link and place it under the dataset folder. The data path should look like dataset/sequence/00. The ground truth poses are provided under the dataset/poses folder. Please modify params.py to set up the right path.

Preprocessing

Run preprocessing.py to get the ground truth for relative poses. The 6DoF relative poses are represented as rotations (roll, pitch, yaw) and translations (x, y, z).

Pre-trained FlowNet model

We refer people to the original github to download the pre-trained FlowNet. Please modify params.py to set up the right path.

Train the DeepVO model

Run train.py to train the DeepVO model. The model will be saved at experiments/experiment_name/models and the log file will be stored under experiments/experiment_name/record. The hyper-parameters are included in params.py and instructions are provided inside. The default image size is set to be 192 x 640

Test the DeepVO model

Run test.py to generate estimations under experiments/experiment_name/results.

To evaluate the test results, run evaluation.py to generate visualizations under experiments/experiment_name/results. The evaluation code is borrowed from link.

Results

03

04

05

06

07

10