Complex YOLOv4

The PyTorch Implementation based on YOLOv4 of the paper: Complex-YOLO: Real-time 3D Object Detection on Point Clouds



2. Getting Started

2.1. Requirement

pip install -U -r requirements.txt

For mayavi and shapely
libraries, please refer to the installation instructions from their official websites.

2.2. Data Preparation

Download the 3D KITTI detection dataset from here.

The downloaded data includes:

  • Velodyne point clouds (29 GB): input data to the Complex-YOLO model
  • Training labels of object data set (5 MB): input label to the Complex-YOLO model
  • Camera calibration matrices of object data set (16 MB): for visualization of predictions
  • Left color images of object data set (12 GB): for visualization of predictions

Please make sure that you construct the source code & dataset directories structure as below.

For 3D point cloud preprocessing, please refer to the previous works:

2.3. Complex-YOLO architecture


This work has been based on YOLOv4 for 2D object detection. Please refer to the original paper of YOLOv4
and the Pytorch implementation which is the great work from Tianxiaomo

2.4. How to run

2.4.1. Visualize the dataset (both BEV images from LiDAR and camera images)

cd src/data_process
  • To visualize BEV maps and camera images (with 3D boxes), let's execute (the output-width param can be changed to
    show the images in a bigger/smaller window)
python --output-width 608
  • To visualize mosaics that are composed from 4 BEV maps (Using during training only), let's execute:
python --show-train-data --mosaic --output-width 608 

By default, there is no padding for the output mosaics, the feature could be activated by executing:

python --show-train-data --mosaic --random-padding --output-width 608 

2.4.2. Inference

python --gpu_idx 0 --pretrained_path <PATH>...

The trained model will be provided soon. Please watch the repo to get notifications for next update.

2.4.3. Training Single machine, single gpu
python --gpu_idx 0 --multiscale_training --batch_size <N> --num_workers <N>... Multi-processing Distributed Data Parallel Training

We should always use the nccl backend for multi-processing distributed training since it currently provides the best
distributed training performance.

  • Single machine (node), multiple GPUs
python --dist-url 'tcp://' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0
  • Two machines (two nodes), multiple GPUs

First machine

python --dist-url 'tcp://IP_OF_NODE1:FREEPORT' --dist-backend 'nccl' --multiprocessing-distributed --world-size 2 --rank 0

Second machine

python --dist-url 'tcp://IP_OF_NODE2:FREEPORT' --dist-backend 'nccl' --multiprocessing-distributed --world-size 2 --rank 1

To reproduce the results, you can run the bash shell script


2.4.4. Evaluation

python ----gpu_idx 0 --pretrained_path <PATH> --img_size <SIZE> --conf-thresh <THRESH> --nms-thresh <THRESH> --iou-thresh <THRESH>...

The comparison of this implementation with Complex-YOLOv2, Complex-YOLOv3 will be updated soon.

mAP Comparison (min 0.50 IoU)

Model/Class Car Pedestrian Cyclist Average

2.5. List of usage for Bag of Freebies (BoF) & Bag of Specials (BoS) in this implementation

Backbone Detector
BoF [x] Dropblock
[x] Random rescale, rotation (global)
[x] Mosaic augmentation
[x] Cross mini-Batch Normalization
[x] Dropblock
[x] Random training shapes
BoS [x] Mish activation
[x] Cross-stage partial connections (CSP)
[x] Multi-input weighted residual connections (MiWRC)
[x] Mish activation
[x] SPP-block
[x] SAM-block
[x] PAN path-aggregation block
[ ] CIoU/GIoU loss


If you think this work is useful, please give me a star!

If you find any errors or have any suggestions, please contact me (Email: [email protected]).

Thank you!


  author = {Martin Simon, Stefan Milz, Karl Amende, Horst-Michael Gross},
  title = {Complex-YOLO: Real-time 3D Object Detection on Point Clouds},
  year = {2018},
  journal = {arXiv},

  author = {Alexey Bochkovskiy, Chien-Yao Wang, Hong-Yuan Mark Liao},
  title = {YOLOv4: Optimal Speed and Accuracy of Object Detection},
  year = {2020},
  journal = {arXiv},

Folder structure

└── dataset/    
    └── kitti/
        │   ├── train.txt
        │   └── val.txt
        ├── training/
        │   ├── image_2/ <-- for visualization
        │   ├── calib/
        │   ├── label_2/
        │   └── velodyne/
        └── testing/  
        │   ├── image_2/ <-- for visualization
        │   ├── calib/
        │   └── velodyne/ 
        └── classes_names.txt
└── src/
    ├── config/
    │   ├── complex_yolov4.cfg
    │   ├── complex_yolov4-tiny.cfg
    │   ├──
    │   └──
    ├── data_process/
    │   ├──
    │   ├──
    │   ├──
    │   ├──
    │   ├──
    │   └──
    ├── models/
    │   ├──
    │   ├──
    │   ├──
    │   ├──
    │   ├──
    │   └──
    └── utils/
    │   ├──
    │   ├──
    │   ├──
    │   ├──
    │   ├──
    │   ├──
    │   ├──
    │   ├──
    │   └──
└── requirements.txt


usage: [-h] [--seed SEED] [--saved_fn FN] [--working-dir PATH]
                [-a ARCH] [--cfgfile PATH] [--pretrained_path PATH]
                [--img_size IMG_SIZE] [--multiscale_training] [--mosaic]
                [--random-padding] [--no-val] [--num_samples NUM_SAMPLES]
                [--num_workers NUM_WORKERS] [--batch_size BATCH_SIZE]
                [--subdivisions SUBDIVISIONS] [--print_freq N]
                [--tensorboard_freq N] [--checkpoint_freq N] [--start_epoch N]
                [--num_epochs N] [--lr LR] [--minimum_lr MIN_LR]
                [--momentum M] [-wd WD] [--optimizer_type OPTIMIZER]
                [--burn_in N] [--steps [STEPS [STEPS ...]]] [--world-size N]
                [--rank N] [--dist-url DIST_URL] [--dist-backend DIST_BACKEND]
                [--gpu_idx GPU_IDX] [--no_cuda]
                [--multiprocessing-distributed] [--evaluate]
                [--resume_path PATH] [--conf-thresh CONF_THRESH]
                [--nms-thresh NMS_THRESH] [--iou-thresh IOU_THRESH]

The Implementation of Complex YOLOv4

optional arguments:
  -h, --help            show this help message and exit
  --seed SEED           re-produce the results with seed random
  --saved_fn FN         The name using for saving logs, models,...
  --working-dir PATH    The ROOT working directory
  -a ARCH, --arch ARCH  The name of the model architecture
  --cfgfile PATH        The path for cfgfile (only for darknet)
  --pretrained_path PATH
                        the path of the pretrained checkpoint
  --img_size IMG_SIZE   the size of input image
                        If true, use scaling data for training
  --mosaic              If true, compose training samples as mosaics
  --random-padding      If true, random padding if using mosaic augmentation
  --no-val              If true, dont evaluate the model on the val set
  --num_samples NUM_SAMPLES
                        Take a subset of the dataset to run and debug
  --num_workers NUM_WORKERS
                        Number of threads for loading data
  --batch_size BATCH_SIZE
                        mini-batch size (default: 4), this is the totalbatch
                        size of all GPUs on the current node when usingData
                        Parallel or Distributed Data Parallel
  --subdivisions SUBDIVISIONS
                        subdivisions during training
  --print_freq N        print frequency (default: 50)
  --tensorboard_freq N  frequency of saving tensorboard (default: 20)
  --checkpoint_freq N   frequency of saving checkpoints (default: 2)
  --start_epoch N       the starting epoch
  --num_epochs N        number of total epochs to run
  --lr LR               initial learning rate
  --minimum_lr MIN_LR   minimum learning rate during training
  --momentum M          momentum
  -wd WD, --weight_decay WD
                        weight decay (default: 1e-6)
  --optimizer_type OPTIMIZER
                        the type of optimizer, it can be sgd or adam
  --burn_in N           number of burn in step
  --steps [STEPS [STEPS ...]]
                        number of burn in step
  --world-size N        number of nodes for distributed training
  --rank N              node rank for distributed training
  --dist-url DIST_URL   url used to set up distributed training
  --dist-backend DIST_BACKEND
                        distributed backend
  --gpu_idx GPU_IDX     GPU index to use.
  --no_cuda             If true, cuda is not used.
                        Use multi-processing distributed training to launch N
                        processes per node, which has N GPUs. This is the
                        fastest way to use PyTorch for either single node or
                        multi node data parallel training
  --evaluate            only evaluate the model, not training
  --resume_path PATH    the path of the resumed checkpoint
  --conf-thresh CONF_THRESH
                        for evaluation - the threshold for class conf
  --nms-thresh NMS_THRESH
                        for evaluation - the threshold for nms
  --iou-thresh IOU_THRESH
                        for evaluation - the threshold for IoU