3D Object detection with Lyft pointnet data

This project adapted the code of frustum-pointnet to solve this past Kaggle competition Lyft 3D Object Detection for Autonomous Vehicles

Main workflow of training and testing

Setting the path configuration

python setup_path.py --model_checkpoint /path/to/model/checkpt --data_path /path/to/data --artifact_path /path/to/artifact --object_detection_model_path /object/detection/model/path

artifact path are the paths that store user-generated data.

Training the point clouds by frustum pointnetV1

Prepare the frustum pointnet data of selected scenes in Lyft train data:

python run_prepare_lyft_data.py --scenes scene_1,scene_2,... --data_type {train, test} [--from_rgb]

- ```--data_type```: sets the type of the data. 
- If ```--from_rgb``` is set, the frustum will be generated by a 2D object detector. If ```--from_rgb``` was not set, the frustum is selected by ground truths.

To preprocess all scenes:

python run_prepare_lyft_data.py --scenes all

Or use scripts prepare_data_from_rgb.sh or prepare_data_from_gt.sh.

Train the model Make model setting in lyft_train_v2_on_local.sh or lyft_train_v2_on_aws.sh. Then execute the scripts.

sh lyft_train_v2_on_local.sh

Run inference with the trained model

sh lyft_test_v2_on_local.sh

So far, the program searches the pattern scene_\d+_train.tfrec in the designated directories assigned in lyft_train_v2_on_local.sh and lyft_test_v2_on_local.sh.

Run infer_train_eval_data.sh
Run test_score_calculator.py

To generate the file for Kaggle submission, run merge_prediction.py

Testing

The workflow of running testing:

For each sample token, run_prepare_lyft_data.py run through the following process

Load camera image
Use a pretrained 2D detector to find the 2D boxes
Cut out the frustums from the pretrained 2D boxes

However, the second step requires a GPU to compute efficiently. To save the cost of running the 2D detector on AWS, I would like a GPU instance focuses on step 2.

The workflow then becomes the following:

Find all the images associated with a sample token in one go. (using CPU): generate_image_file.py
Run the 2D detector (using a GPU instance) detect_all_images.py
Cut out frustums. (using CPU) run_prepare_lyft_data.py --use_detected_2d with

With step 1 executed in a CPU instance, it can saves the 50% of the time on a GPU instance, compared to running step 1 and step 2 in a GPU instance.

Not completed yet:

Use the function parse_pointnet_output.get_box_from_inference() to transform the inferred results back to world coordinates. This needs two steps:

Correct the frustum angle:

The predicted heading angle
The predicted center

Transform the predicted corners from camera coordinates to world coordinates.

Training 2D object detector

Install Tensorflow object detection API

TODO: installation instruction.

Prepare the training data

This 2D object detector uses Tensorflow object detection API. It was tested on tensorflow 1.14, and does not support Tensorflow 2.0 yet.

Prepare the object detection data to match the format for Tensorflow object detection API. In the last line of prepare_object_detection_data.py, change write_data_to_files(param). param is the number of sample tokens in train.csv to be run through. Set param=None to go through all sample tokens.

python prepare_object_detection_data.py.

Train the model

The model settings are in ./object_detection_models/models. Configure the model in XXXXXX.config, run_model.sh, and then run run_model.sh. The category file is object_detecton_models/models/lyft_object_map.pbtxt.

Export the model

Run export_model.sh. TODO: the path setting somehow does not allow the model to be used on local machine, this may due to that the path setting in the configuration files are absolute rather than relative.

Run inference

TODO: Full scripts to detect all data not completed yet.

Set the object detection model path in user_config.txt. See default_config.txt for examples.
See parse_pointnet_output.py for an example of running object detection

Visualization

Bird view frustums: [Visualize frustum pipelines.ipynb](Visualize frustum pipelines.ipynb)

Visualize predicted results in 3D: pred_viewer_test.py

Plot frustum in 3D: plot_v2_data.py

Test plot frustums and rotated frustums point cloud points: prepare_lyft_data_v2_test.test_plot_one_frustum

Test plot frustums and rotated frustums point cloud points with rgb detection: prepare_lyft_data_v2_rgb_test.FrustumRGBTestCase.test_plot_frustums

Visualize predicted results in 3D from RGB data: view_full_pipeline.plot_prediction_data (work in progress)

kanhua/lyft-3d-main