百科.dev
全部条目在线工具AI 编程趋势榜开源项目技术资讯提交条目
登录
< 返回工具列表
C

Complex-YOLOv3

> 编程语言
开源

Complex-YOLO-v3: 实时 3D 对象检测 ✅ PyTorch 实现

200 stars0 点赞0 次浏览
访问官网GitHub

工具介绍

Complex-YOLO-v3: 实时 3D 对象检测 ✅ PyTorch 实现

Complex-YOLO-V3

Complete but Unofficial PyTorch Implementation of Complex-YOLO: Real-time 3D Object Detection on Point Clouds with YoloV3

Installation

Clone the project and install requirements

$ git clone https://github.com/ghimiredhikura/Complex-YOLOv3
$ cd Complex-YOLO-V3/
$ sudo pip install -r requirements.txt

Quickstart

Download pretrained weights [yolov3, tiny-yolov3]

$ cd checkpoints/
$ python download_weights.py

Test [without downloading dataset]

1. $ python test_detection.py --split=sample --folder=sampledata  
2. $ python test_both_side_detection.py --split=sample --folder=sampledata

Demo Video [Click to Play]

complex-yolov3_gif

Data Preparation

You can see sampledata folder in data/KITTI/dataset directory which can be used for testing this project without downloading KITTI dataset. However, if you want to train the model by yourself and check the mAP in validation set just follow the steps below.

Download the 3D KITTI detection dataset

  1. Camera calibration matrices of object data set (16 MB)
  2. Training labels of object data set (5 MB)
  3. Velodyne point clouds (29 GB)
  4. Left color images of object data set (12 GB)

Now you have to manage dataset directory structure. Place your dataset into data folder. Please make sure that you have the dataset directory structure as follows.

└── data/KITTI/object
       ├── training    <-- 7481 train data
       |   ├── image_2 <-- for visualization
       |   ├── calib
       |   ├── label_2
       |   ├── velodyne
       └── testing     <-- 7580 test data
           ├── image_2 <-- for visualization
           ├── calib
           ├── velodyne           

The train/valid split of training dataset as well as sample and test dataset ids are in data/KITTI/ImageSets directory. From training set of 7481 images, 6000 images are used for training and remaining 1481 images are used for validation. The mAP results reported in this project are evaluated into this valid set with custom mAP evaluation script with 0.5 iou for each object class.

Verify your download

$ python check_dataset.py

Train

$ train.py [-h] [--epochs EPOCHS] [--batch_size BATCH_SIZE]
            [--gradient_accumulations GRADIENT_ACCUMULATIONS]
            [--model_def MODEL_DEF] 
            [--pretrained_weights PRETRAINED_WEIGHTS] 
            [--n_cpu N_CPU] [--img_size IMG_SIZE]
            [--evaluation_interval EVALUATION_INTERVAL]
            [--multiscale_training MULTISCALE_TRAINING]

--Training log example--

---- [Epoch 0/300, Batch 250/1441] ----  
+------------+--------------+--------------+--------------+  
| Metrics    | YOLO Layer 0 | YOLO Layer 1 | YOLO Layer 2 |  
+------------+--------------+--------------+--------------+  
| grid_size  | 17           | 34           | 68           |  
| loss       | 6.952686     | 5.046788     | 4.256296     |  
| x          | 0.054503     | 0.047048     | 0.060234     |  
| y          | 0.110871     | 0.059848     | 0.081368     |
| w          | 0.101059     | 0.056696     | 0.022349     |
| h          | 0.294365     | 0.230845     | 0.076873     |
| im         | 0.215230     | 0.218564     | 0.184226     |
| re         | 1.049812     | 0.883522     | 0.783887     |
| conf       | 4.682138     | 3.265709     | 2.941420     |
| cls        | 0.444707     | 0.284557     | 0.105938     |
| cls_acc    | 67.74%       | 83.87%       | 96.77%       |
| recall50   | 0.000000     | 0.129032     | 0.322581     |
| recall75   | 0.000000     | 0.032258     | 0.032258     |
| precision  | 0.000000     | 0.285714     | 0.133333     |
| conf_obj   | 0.058708     | 0.248192     | 0.347815     |
| conf_noobj | 0.014188     | 0.007680     | 0.010709     |
+------------+--------------+--------------+--------------+
Total loss 16.255769729614258
---- ETA 0:18:27.490254

Test

$ python test_detection.py
$ python test_both_side_detection.py

Evaluation

$ python eval_mAP.py 

mAP (min. 50 IoU)

Model/Class Car Pedestrian Cyclist Average
Complex-YOLO-v3 97.89 82.71 90.12 90.24
Complex-Tiny-YOLO-v3 95.91 49.29 78.75 74.65

Results

Credit

  1. Complex-YOLO: https://arxiv.org/pdf/1803.06199.pdf

YoloV3 Implementation is borrowed from:

  1. https://github.com/eriklindernoren/PyTorch-YOLOv3

Point Cloud Preprocessing is based on:

  1. https://github.com/skyhehe123/VoxelNet-pytorch
  2. https://github.com/dongwoohhh/MV3D-Pytorch

Issues· 52 开放

查看全部 Issues在 GitHub 打开

暂无开放 Issues,或尚未同步最近议题。

> 标签

Pythondeep-learningobject-detectionpytorch

暂无评论,来聊聊你的看法吧

> 工具信息

发布日期2026年8月1日
最后更新2026年9月18日
分类编程语言
定价开源

> 相关工具

T
TypeScript
JavaScript 的超集,为前端与全栈提供静态类型
P
Python
通用编程语言,广泛用于 Web、数据与 AI
G
Go
Google 推出的简洁高效系统语言