DeLS-3D

The code for DeLS-3D of CVPR 2018 and the jounal paper, website: Apolloscape Dataset.

    @inproceedings{wang2018dels,
        title={DeLS-3D: Deep Localization and Segmentation with a 3D Semantic Map},
        author={Wang, Peng and Yang, Ruigang and Cao, Binbin and Xu, Wei and Lin, Yuanqing},
        booktitle={CVPR},
        pages={5860--5869},
        year={2018}
    }

    @article{huangwang2018_apolloscape,
        author    = {Xinyu Huang and
                   Peng Wang and
                   Xinjing Chen and
                   Qichuan Geng and
                   Dingfu Zhou and
                   Ruigang Yang},
        title     = {The ApolloScape Open Dataset for Autonomous Driving and its Application},
        journal   = {CoRR},
        volume    = {abs/1803.06184},
        year      = {2018}
    }

Log:

• zpark testing code is ready to use.
• Dataset of zpark & dlake are availabe

Dependents

Our system depends on:

API that supporting rendering of 3D points to label maps and depth maps: apolloscape-api.

Various API that supporting many 3D applications vis_utils

Tested with Ubuntu 14.04 for rendering , python version 2.7.5, mxnet 0.11 for training.

If you need to train: We use the code for augmentation: imgaug

Dataset

Each part of the dataset are including, images, camera_pose, semantic_label, semantic_3D_points, split.

Results videos:

Zpark

Dlake

Notice Dlake data is a subset of official released data road01_inst.tar.gz where you may obtain corresponding instance segmentation, depth and poses also. We do different split for train/val used in the paper.

In addition, for semantic 3D points, we release a merged pcd file which is used for rendering label maps in the paper, we also have much denser point cloud that are separated stored for Dlake dataset.

[comment]: # Download and unzip to folder data, perturbed poses are also provided for results reproducibility. Download and unzip to folder data.

Testing

Notice the number of the pre-trained models could be slightly different than that from paper due to the randomness from perturbation of GPS/IMU, but it should be close.

source.rc Firstly, pull the dependents and source the corresponding root folders

Download pre-trained models

Zpark: Download

Dlake: Download

Download them and put under 'models/${Data}/' ${Data} is the corresponding dataset name.

Testing

Currently only zpark testing code is available

run the following code for test and evaluation of pose_cnn, pose_rnn, and seg_cnn.

- the images has some area (face&plate) blurred due to security reasons, therefore the number output from our pretrained models are slightly different from that in paper. You may need to retrain the model for better performance.

- The test code does not have clipping pose inside road area as preprocessing indicated due to liscense issue

Download

put under 'results/${Data}/noisy_pose' You may also use the code to generate simulated noisy pose too.

Set up the environment

source source.rc

python test_DeLS-3D.py --dataset zpark --pose_cnn ./models/zpark/pose_cnn-0000 --pose_rnn models/zpark/pose_rnn-0000 --seg_cnn models/zpark/seg_cnn-0000

Should get results close to the paper with a random pose simulation. Notice current pipeline is an offline version, since CNN and RNN are not connected, one may need to reimplement for online version.

Training

Some loss functions are uploaded but the pipeline need to be cleaned. To be updated.

Note

I may not have enough time for solving all your issues, expect delay of reply.

Contact: wangpeng54@baidu.com