Home

Awesome

MonoPSR

Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction

Jason Ku*, Alex D. Pon*, Steven L. Waslander (*Equal Contribution)

This repository contains the public release of the Tensorflow implementation of Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction in CVPR 2019.

Video

Demo video showing results on several KITTI sequences.

Demo

Getting Started

Implemented and tested on Ubuntu 16.04 with Python 3.5 and Tensorflow 1.8.0.

Clone this repo

git clone git@github.com:kujason/monopsr.git

Install Python dependencies

cd monopsr
pip3 install -r requirements.txt

Add monopsr/src to your PYTHONPATH

# For virtualenvwrapper users
add2virtualenv src/.

Compile the two custom TF ops src/tf_ops/nn_distance and src/tf_ops/approxmatch by running the shell scripts found in the respective folders. The location of your TensorFlow python package is passed as an argument.

For example:

sh src/tf_ops/approxmatch/tf_approxmatch_compile.sh ${HOME}/.virtualenvs/{monopsr}/lib/python3.5/site-packages/tensorflow

sh src/tf_ops/nn_distance/tf_nndistance_compile.sh ${HOME}/.virtualenvs/{monopsr}/lib/python3.5/site-packages/tensorflow

Training

To train on the KITTI Object Detection Dataset:

Download the data and place it in your home folder at ~/Kitti/object.

Go here and download the train.txt, val.txt and trainval.txt splits into ~/Kitti/object.

/home/$USER/Kitti
    object
        testing
        training
            calib
            image_2
            label_2
            velodyne
        train.txt
        trainval.txt
        val.txt

2D Detections

Download the MSCNN 2D detections here and place it in monopsr/data/detections/mscnn

Depth Maps and Instance Masks

Generate the ground truth depth maps and instance segmentation:

python demos/depth_completion/save_lidar_depth_maps.py
python demos/instances/gen_instance_masks.py

Place the depth maps and segmentation outputs in ~/Kitti/object/training/.

/home/$USER/Kitti
    object
        testing
        training
            calib
            *depth_2_multiscale
            image_2
            *instance_2_multiscale
            label_2
            velodyne
        train.txt
        val.txt

* denotes generated folders

Pre-trained ResNet-101

Download the pre-trained ResNet-101 model (faster_rcnn_resnet101_kitti) from the Tensorflow Object Detection API here, and extract it in data/pretrained as data/pretrained/faster_rcnn_resnet101_kitti_2018_01_28

Model Configuration

A sample configuration for training is in src/monopsr/configs. You can train using the example configs, or modify an existing configuration.

Run Training

To start training, run the following:

python src/monopsr/experiments/run_training.py --config_path='src/monopsr/configs/monopsr_model_000.yaml'

Run Evaluation

To start evaluation, run the following:

python src/monopsr/experiments/run_evaluation.py --config_path='src/monopsr/configs/monopsr_model_000.yaml'

Note, we primarily use this script to determine metrics on the centroid and point cloud estimation. This is not used to obtain the validation results in the paper since it uses some ground truth boxes. To get the validation results in the paper we use run_inference.py.

Run Inference

To start inference, run the following:

python src/monopsr/experiments/run_inference.py --config_path='src/monopsr/configs/monopsr_model_000
.yaml' --default_ckpt_num='100000' --data_split='val'

To calculate AP performance, follow the instructions in scripts/offline_eval/kitti_native_eval

Contact

Please contact kujason.ku@mail.utoronto.ca or alex.pon@mail.utoronto.ca for any questions or issues.