Pytorch-Yolo-3d-Yolov3

English Tutorial : https://wikidocs.net/167908

한글 Tutorial : https://wikidocs.net/163178

Pytorch-Yolo-3d-Yolov3

Complete but Unofficial PyTorch Implementation of YOLO3D: End-to-end real-time 3D Oriented Object Bounding Box Detection from LiDAR Point Cloud (ECCV 2018) with YoloV3 Code was tested with following specs:

Code was tested on Windows 10

1. Installation

First, clone or download this GitHub repository. Install requirements and download from official darknet weights:

# yolov3
wget -P model_data https://pjreddie.com/media/files/yolov3.weights

# yolov3-tiny
wget -P model_data https://pjreddie.com/media/files/yolov3-tiny.weights

# yolov4
wget -P model_data https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.weights

# yolov4-tiny
wget -P model_data https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights

Or you can download darknet weights from my google drive:

https://drive.google.com/drive/folders/1w4KNO2jIlkyzQgUkkcZ18zrC8A1Nqvwa?usp=sharing

2. Pretrained weights

You can download darknet weights from my google drive:

Yolo3d - Yolo v3:

https://drive.google.com/file/d/15Hza0MjND3EWJDOt8R0XBXOZnCRoMA2o/view?usp=sharing

Yolo3d - Yolo v3 - tiny:

https://drive.google.com/file/d/1IUXbc2dIm-Qowhwb44aCHJze536dJcwU/view?usp=sharing

3. Quick start

3.1. Download pretrained weights

  ${ROOT}
  ├── checkpoints/
  │    ├── Model_yolo3d_yolov3.pth
  │    ├── Model_yolo3d_yolov3_tiny.pth
  │    ├── yolov3.weights
  │    └── yolov3-tiny.weights

3.2. Dataset shall contains shall contains belows.

detect_1 folder shall contain folders and contents.

ImageSets folder shall contain *.txt files.

${ROOT}
├── dataset/
│    ├── classes.names
│    └── kitti/
│          ├── classes_names.txt
│          ├── detect_1/    <-- for detection test
│          │    ├── calib/
│          │    ├── image_2/
│          │    └── velodyne/
│          ├── ImageSets/ 
│          │    ├── detect_1.txt
│          │    ├── detect_2.txt
│          │    ├── sample.txt
│          ├── sampledata/ 
│          │    ├── image_2/
│          │    ├── calib/
│          │    ├── label_2/
│          │    └── velodyne/

3.3 Test [without downloading dataset]

Detection test for Yolo v3 with detect_1 folder.

 $ python detection.py --model_def config/cfg/yolo3d_yolov3.cfg --pretrained_path checkpoints/Model_yolo3d_yolov3.pth
 or       
 $ python detection.py

Detection test for Yolo v3-tiny with detect_1 folder.

 $ python detection.py --model_def config/cfg/yolo3d_yolov3_tiny.cfg --pretrained_path checkpoints/Model_yolo3d_yolov3_tiny.pth

3.4 Demo Video

One side detection demo.

4. Data Preparation from KITTI

You can see sampledata folder in dataset/kitti/sampledata 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

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

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

The train/valid split of training dataset as well as sample and test dataset ids are in dataset/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.

5. Data Preparation from my google drive

If you download the dataset from the KITTI Vision Benchmark Suite, it is difficult to build detect dataset. So, I recommend you to download dataset from my google drive.

https://drive.google.com/file/d/12Fbtxo2F8Nn9dHDV9UPnovVaCf-Uc89R/view?usp=sharing

5.1 Dataset folder structure shall be

${ROOT}
├── dataset/
│    ├── classes.names
│    └── kitti/
│          ├── classes_names.txt
│          ├── detect_1/    <-- for detection test
│          │    ├── calib/
│          │    ├── image_2/
│          │    └── velodyne/
│          ├── detect_2/    <-- for detection test
│          │    ├── calib/
│          │    ├── image_2/
│          │    └── velodyne/
│          ├── ImageSets/ 
│          │    ├── detect_1.txt
│          │    ├── detect_2.txt
│          │    ├── sample.txt
│          │    ├── test.txt
│          │    ├── train.txt
│          │    ├── val.txt
│          │    └── valid.txt
│          ├── sampledata/ 
│          │    ├── image_2/
│          │    ├── calib/
│          │    ├── label_2/
│          │    └── velodyne/
│          ├── training/    <-- 7481 train data
│          │    ├── image_2/  <-- for visualization
│          │    ├── calib/
│          │    ├── label_2/
│          │    └── velodyne/
│          └── testing/     <-- 7580 test data
│                ├── image_2/  <-- for visualization
│                ├── calib/
│                └── velodyne/

6. Train

Yolo3d - Yolo v3 training from pretrained weight.

 $ python train.py --model_def config/cfg/yolo3d_yolov3.cfg --pretrained_path checkpoints/Model_yolo3d_yolov3.pth --save_path checkpoints/Model_yolo3d_yolov3.pth
 
 or
 
 $ python train.py

Yolo3d - Yolo v3 training from darknet weight.

 $ python train.py --model_def config/cfg/yolo3d_yolov3.cfg --pretrained_path checkpoints/yolov3.weights --save_path checkpoints/Model_yolo3d_yolov3.pth

Yolo3d - Yolo v3-tiny training from pretrained weight.

 $ python train.py  --model_def config/cfg/yolo3d_yolov3_tiny.cfg --pretrained_path checkpoints/Model_yolo3d_yolov3_tiny.pth --save_path checkpoints/Model_yolo3d_yolov3_tiny.pth

Yolo3d - Yolo v3-tiny training from darknet weight.

 $ python train.py  --model_def config/cfg/yolo3d_yolov3_tiny.cfg --pretrained_path checkpoints/yolov3-tiny.weights --save_path checkpoints/Model_yolo3d_yolov3_tiny.pth

7. Evaluation

Yolo3d - Yolo v3 evaluation.

 $ python eval_mAP.py 
 
 or
 
 $ python eval_mAP.py --model_def config/cfg/yolo3d_yolov3.cfg --pretrained_path checkpoints/Model_yolo3d_yolov3.pth

      100%|██████████████████████████████████████████████████████████████████████████| 354/354 [02:50<00:00,  2.07it/s]
      Computing AP: 100%|███████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 111.53it/s]

      Done computing mAP...

        >>>      Class 0 (Car): precision = 0.9115, recall = 0.9753, AP = 0.9688, f1: 0.9423
        >>>      Class 1 (Ped): precision = 0.6961, recall = 0.9306, AP = 0.7848, f1: 0.7964
        >>>      Class 2 (Cyc): precision = 0.8000, recall = 0.9377, AP = 0.9096, f1: 0.8634

      mAP: 0.8877

Yolo3d - Yolo v3 - tiny evaluation.

 $ python eval_mAP.py --model_def config/cfg/yolo3d_yolov3_tiny.cfg --pretrained_path checkpoints/Model_yolo3d_yolov3_tiny.pth

      100%|████████████████████████████████████████████████████████████████████████████| 354/354 [01:37<00:00,  3.63it/s]
      Computing AP: 100%|█████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 107.29it/s]

      Done computing mAP...

        >>>      Class 0 (Car): precision = 0.6226, recall = 0.9699, AP = 0.7933, f1: 0.7584
        >>>      Class 1 (Ped): precision = 0.1332, recall = 0.3299, AP = 0.0548, f1: 0.1898
        >>>      Class 2 (Cyc): precision = 0.1299, recall = 0.3626, AP = 0.0663, f1: 0.1913

      mAP: 0.3048

Credit

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

YoloV3 Implementation is borrowed from:

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

Point Cloud Preprocessing is based on:

Folder structure

${ROOT}
├── detection.py
├── eval_mAP.py
├── README.md 
├── train.py
├── checkpoints/ 
│    ├── Model_yolo3d_yolov3.pth
│    ├── Model_yolo3d_yolov3_tiny.pth
│    ├── yolov3.weights
│    └── yolov3-tiny.weights
├── config/ 
│    ├── cfg/
│    │     ├── yolo3d_yolov3.cfg
│    │     ├── yolo3d_yolov3_tiny.cfg
│    │     ├── yolo3d_yolov4.cfg
│    │     └── yolo3d_yolov4_tiny.cfg
│    ├── kitti_config.py
│    └── train_config.py
├── data_process/ 
│    ├── kitti_aug_utils.py
│    ├── kitti_bev_utils.py
│    ├── kitti_dataloader.py
│    ├── kitti_dataset.py
│    └── kitti_utils.py
├── dataset/
│    ├── classes.names
│    └── kitti/
│          ├── classes_names.txt
│          ├── detect_1/    <-- for detection test
│          │    ├── calib/
│          │    ├── image_2/
│          │    └── velodyne/
│          ├── detect_2/    <-- for detection test
│          │    ├── calib/
│          │    ├── image_2/
│          │    └── velodyne/
│          ├── ImageSets/ 
│          │    ├── detect_1.txt
│          │    ├── detect_2.txt
│          │    ├── sample.txt
│          │    ├── test.txt
│          │    ├── train.txt
│          │    ├── val.txt
│          │    └── valid.txt
│          ├── sampledata/ 
│          │    ├── image_2/
│          │    ├── calib/
│          │    ├── label_2/
│          │    └── velodyne/
│          ├── training/    <-- 7481 train data
│          │    ├── image_2/  <-- for visualization
│          │    ├── calib/
│          │    ├── label_2/
│          │    └── velodyne/
│          └── testing/     <-- 7580 test data
│                ├── image_2/  <-- for visualization
│                ├── calib/
│                └── velodyne/
├── logs/ 
├── models/
│    ├── darknet_utils.py
│    ├── darknet2pytorch.py
│    ├── model_utils.py
│    ├── models.py
│    └── yolo_layer.py
└── utils/ 
      ├── evaluation_utils.py
      ├── find_anchors.py
      ├── iou_rotated_boxes_utils.py
      ├── mayavi_viewer.py
      ├── misc.py
      ├── torch_utils.py
      ├── train_utils.py
      └── utils.py

Name		Name	Last commit message	Last commit date
Latest commit History 84 Commits
asset		asset
checkpoints		checkpoints
config		config
data_process		data_process
dataset		dataset
logs		logs
models		models
utils		utils
How_to_train_YOLO_3d_Yolo_V3.ipynb		How_to_train_YOLO_3d_Yolo_V3.ipynb
Notebook_Pytorch_Yolo_3d_Yolov3.ipynb		Notebook_Pytorch_Yolo_3d_Yolov3.ipynb
README.md		README.md
detection.py		detection.py
eval_mAP.py		eval_mAP.py
train.py		train.py
train_old.py		train_old.py

RichardMinsooGo-ML/Bible_3_41_Pytorch_Yolo_3d_Yolov3

Folders and files

Latest commit

History

Repository files navigation

English Tutorial : https://wikidocs.net/167908

한글 Tutorial : https://wikidocs.net/163178

Pytorch-Yolo-3d-Yolov3

1. Installation

2. Pretrained weights

3. Quick start

3.1. Download pretrained weights

3.2. Dataset shall contains shall contains belows.

3.3 Test [without downloading dataset]

3.4 Demo Video

4. Data Preparation from KITTI

Download the 3D KITTI detection dataset

5. Data Preparation from my google drive

5.1 Dataset folder structure shall be

6. Train

7. Evaluation

Credit

Folder structure

About

Resources

Stars

Watchers

Forks

Languages