CLOCs is a novel Camera-LiDAR Object Candidates fusion network. It provides a low-complexity multi-modal fusion framework that improves the performance of single-modality detectors. CLOCs operates on the combined output candidates of any 3D and any 2D detector, and is trained to produce more accurate 3D and 2D detection results.
Tested on python3.6, pytorch 1.1.0, Ubuntu 16.04/18.04.
new 40 recall points
Car: [email protected] [email protected] [email protected]
bev: AP: 96.51 / 95.61, 92.37 / 89.54, 89.41 / 86.96
3d: AP: 92.74 / 90.97, 82.90 / 79.94, 77.75 / 77.09
old 11 recall points
Car: [email protected] [email protected] [email protected]
bev: AP: 90.52 / 90.36, 89.29 / 88.10, 87.84 / 86.80
3d: AP: 89.49 / 88.31, 79.31 / 77.99, 77.36 / 76.52
- Support fusion for pedestrian and cyclists on KITTI.
- Support easier testing for other 2D and 3D detectors, clean the code, remove unrelated environment setups to make the code easier to use.
The code is developed based on SECOND-1.5, please follow the SECOND-1.5 to setup the environment, the dependences for SECOND-1.5 are needed.
pip install shapely fire pybind11 tensorboardX protobuf scikit-image numba pillowFollow the instructions to install spconv v1.0 (commit 8da6f96). Although CLOCs fusion does not need spconv, but SECOND codebase expects it to be correctly configured.
Then adding the CLOCs directory to your PYTHONPATH, you could add the following line (change '/dir/to/your/CLOCs/' according to your CLOCs directory) in your .bashrc under home directory.
export PYTHONPATH=$PYTHONPATH:'/dir/to/your/CLOCs/'Download KITTI dataset and organize the files as follows:
└── KITTI_DATASET_ROOT
├── training <-- 7481 train data
| ├── image_2 <-- for visualization
| ├── calib
| ├── label_2
| ├── velodyne
| └── velodyne_reduced <-- empty directory
└── testing <-- 7518 test data
| ├── image_2 <-- for visualization
| ├── calib
| ├── velodyne
| └── velodyne_reduced <-- empty directory
└── kitti_dbinfos_train.pkl
├── kitti_infos_train.pkl
├── kitti_infos_test.pkl
├── kitti_infos_val.pkl
└── kitti_infos_trainval.pkl
Next, you could follow the SECOND-1.5 instructions to create kitti infos, reduced point cloud and groundtruth-database infos, or just download these files from here and put them in the correct directories as shown above.
CLOCs operates on the combined output of a 3D detector and a 2D detector. For this example, we use SECOND as the 3D detector, Cascade-RCNN as the 2D detector.
-
For this example, we use detections with sigmoid scores, you could download the Cascade-RCNN detections for the KITTI train and validations set from here file name:'cascade_rcnn_sigmoid_data', or you could run the 2D detector by your self and save the results for the fusion. You could also use your own 2D detector to generate these 2D detections and save them in KITTI format for fusion.
-
Then download the pretrained SECOND models from here file name: 'second_model.zip', create an empty directory named
model_dirunder your CLOCs root directory and unzip the files tomodel_dir. Your CLOCs directory should look like this:
└── CLOCs
├── d2_detection_data <-- 2D detection candidates data
├── model_dir <-- SECOND pretrained weights extracted from 'second_model.zip'
├── second
├── torchplus
├── README.md
- Then modify the config file carefully:
train_input_reader: {
...
database_sampler {
database_info_path: "/dir/to/your/kitti_dbinfos_train.pkl"
...
}
kitti_info_path: "/dir/to/your/kitti_infos_train.pkl"
kitti_root_path: "/dir/to/your/KITTI_DATASET_ROOT"
}
...
train_config: {
...
detection_2d_path: "/dir/to/2d_detection/data"
}
...
eval_input_reader: {
...
kitti_info_path: "/dir/to/your/kitti_infos_val.pkl"
kitti_root_path: "/dir/to/your/KITTI_DATASET_ROOT"
}
python ./pytorch/train.py train --config_path=./configs/car.fhd.config --model_dir=/dir/to/your_model_dirThe trained models and related information will be saved in '/dir/to/your_model_dir'
File "./pytorch/train.py", line 869, in predict_v2
opp_labels = (box_preds[..., -1] > 0) ^ dir_labels.byte()
RuntimeError: result type Byte can't be cast to the desired output type BoolSolution: change opp_labels = (box_preds[..., -1] > 0) ^ dir_labels.byte() into opp_labels = (box_preds[..., -1] > 0) ^ dir_labels.to(torch.bool). This is because SECOND-V1.5 is written in older pytorch and you have a newer version.
If you have too much NumbaWarning output messages during training/inferece that looks annoying, adding the following code at the beginning of train.py to ignore them:
from numba.core.errors import NumbaDeprecationWarning, NumbaPendingDeprecationWarning,NumbaPerformanceWarning,NumbaWarning
import warnings
warnings.simplefilter('ignore', category=NumbaDeprecationWarning)
warnings.simplefilter('ignore', category=NumbaPendingDeprecationWarning)
warnings.simplefilter('ignore', category=NumbaPerformanceWarning)
warnings.simplefilter('ignore', category=NumbaWarning)
warnings.simplefilter('ignore')
warnings.filterwarnings('ignore')python ./pytorch/train.py evaluate --config_path=./configs/car.fhd.config --model_dir=/dir/to/your/trained_model --measure_time=True --batch_size=1For example if you want to test the pretrained model downloaded from here file name: 'CLOCs_SecCas_pretrained.zip', unzip it, then you could run:
python ./pytorch/train.py evaluate --config_path=./configs/car.fhd.config --model_dir=/dir/to/your/CLOCs_SecCas_pretrained --measure_time=True --batch_size=1If you want to export KITTI format label files, add pickle_result=False at the end of the above commamd.
Step 1: Prepare the 2D detection candidates, run your 2D detector and save the results in KITTI format. It is recommended to run inference with NMS score threshold equals to 0 (no score thresholding), but if you don't know how to setup this, it is also fine for CLOCs.
Step 2: Prepare the 3D detection candidates, run your 3D detector and save the results in the format that SECOND could read, including a matrix with shape of N by 7 that contains the N 3D bounding boxes, and a N-element vector for the 3D confidence scores. 7 parameters correspond to the representation of a 3D bounding box. Be careful with the order and coordinate of the 7 parameters, if the parameters are in LiDAR coordinate, the order should be x, y, z, width, length, height, heading; if the parameters are in camera coordinate, the orderr should be x, y, z, lenght, height, width, heading. The details of the transformation functions can be found in file './second/pytorch/core/box_torch_ops.py'.
Step 3: Since the number of detection candidates are different for different 2D/3D detectors, you need to modify the corresponding parameters in the CLOCs code. Then train the CLOCs fusion. For example, there are 70400 (200x176x2) detection candidates in each frame from SECOND with batch size equals to 1. It is a very large number because SECOND is a one-stage detector, for other multi-stage detectors, you could just take the detection candidates before the final NMS function, that would reduce the number of detection candidates to hundreds or thousands.
Step 4: The output of CLOCs are fused confidence scores for all the 3D detection candidates, so you need to replace the old confidence scores (from your 3D detector) with the new fused confidence scores from CLOCs for post processing and evaluation. Then these 3D detection candidates with the corresponding CLOCs fused scores are treated as the input for your 3D detector post processing functions to generate final predictions for evaluation.
If you find this work useful in your research, please consider citing:
@article{pang2020clocs,
title={CLOCs: Camera-LiDAR Object Candidates Fusion for 3D Object Detection},
author={Pang, Su and Morris, Daniel and Radha, Hayder},
booktitle={2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year={2020}
organization={IEEE}
}
Our code are mainly based on SECOND, thanks for their excellent work!