https://github.com/devbruce/yolov3-tf2
YOLOv3 implementation with TensorFlow2
https://github.com/devbruce/yolov3-tf2
coco evaluation object-detection tensorflow2 tf2 yolo yolov3 yolov3-tiny
Last synced: about 1 month ago
JSON representation
YOLOv3 implementation with TensorFlow2
- Host: GitHub
- URL: https://github.com/devbruce/yolov3-tf2
- Owner: devbruce
- License: mit
- Created: 2021-06-08T07:16:21.000Z (almost 5 years ago)
- Default Branch: master
- Last Pushed: 2021-08-25T06:13:04.000Z (almost 5 years ago)
- Last Synced: 2026-01-02T06:12:57.222Z (5 months ago)
- Topics: coco, evaluation, object-detection, tensorflow2, tf2, yolo, yolov3, yolov3-tiny
- Language: Jupyter Notebook
- Homepage:
- Size: 1.36 MB
- Stars: 2
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# YOLOv3 with TensorFlow2
## Preview
## Performance
- Inference Parameters: (`input_size`: 416, `conf_thr`: 0.05, `nms_iou_thr`: 0.45)
### YOLOv3
#### COCO Evaluation (with 2017 COCO Validation Dataset)
```
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.366
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.640
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.380
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.185
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.403
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.544
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.290
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.425
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.439
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.238
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.477
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.625
```
#### Inference Speed
- GPU(GeForce RTX 3090): About 15 ~ 16 FPS
- CPU(AMD Ryzen 5 5600X 6-Core Processor): About 3 ~ 4 FPS
### YOLOv3 Tiny
#### COCO Evaluation (with 2017 COCO Validation Dataset)
```
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.089
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.176
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.082
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.045
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.261
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.106
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.156
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.160
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.003
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.117
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.367
```
#### Inference Speed
- GPU(GeForce RTX 3090): About 53 ~ 55 FPS
- CPU(AMD Ryzen 5 5600X 6-Core Processor): About 19 ~ 20 FPS
## Set Environment with Docker
### Build Docker Image
```bash
$ docker build -t ${NAME}:${TAG} .
```
### Create Container
Create container with `docker create` or `docker run`
## Initial Settings
Set [`./configs/initial_settings.json`](./configs/initial_settings.json) (Default settings: COCO)
If you want to only inference with uploaded pretrained coco weight file, keep default `classes`.
```python
{
"project_name": "project_name", # Your project name
"model": "yolo_v3", # [yolo_v3, yolo_v3_tiny]
"classes": { # Class index starts from 1 (Not 0)
"1": "class 1",
"2": "class 2",
...
}
}
```
### Set basic checkpoint files (COCO Pretrained Weight Files)
- `./ckpts/yolo_v3_coco.h5` ([Google Drive Link](https://drive.google.com/file/d/1Fp4a42c2bOpDMK9FRgtMJEJ6IqgKp3TX/view?usp=sharing))
- `./ckpts/yolo_v3_tiny_coco.h5` ([Google Drive Link](https://drive.google.com/file/d/1Am96KN-dxZIQKp-t7Z-mquWpu4Ux4cus/view?usp=sharing))
## Inference
> Inference and Visualization Tutorial: [`./tutorial.ipynb`](./tutorial.ipynb)
```python
import cv2
from libs.inference import YoloInf
# Load a image to inference
img_path = '...'
img_arr = cv2.cvtColor(cv2.imread(img_path), cv2.COLOR_BGR2RGB)
# Define YOLO Model to inference
ckpt_path = '...'
yolo_inf = YoloInf(ckpt_path=ckpt_path)
# Get inference result
preds = yolo_inf.get(img_arr=img_arr, conf_thr=0.3) # List of dicts
```
### Inference Output(`preds`) Format
```python
[
{
'bbox': [left:int, top:int, right:int, bottom:int] # [x_min, y_min, x_max, y_max],
'confidence': confidence:float,
'class_index': class_index:int,
'class_name': class_name:str,
},
# ...
]
```
## Training Custom Dataset
1. Set [`./configs/initial_settings.json`](./configs/initial_settings.json) (Compatible with COCO Format GT File)
2. Save dataset as `Dataset Directory Structure` described below
3. Run [`./train.py`](./train.py)
### Training Script
```bash
python train.py
```
#### Options
- `--epochs`: Number of training epochs (Default: `./configs/base.py`)
- `--init_lr`: Initials (Default: `./configs/base.py`)
- `--end_lr`: End learning rate (Default: `./configs/base.py`)
- `--warmup_epochs`: Warm-up epochs (Default: `./configs/base.py`)
- `--batch_size`: Number of batch size (Default: `./configs/base.py`)
- `--transfer_coco`: Transfer pretrained coco weights (Default: `./configs/base.py`)
- `--validation`: Number of training epochs (Default: `True`)
### Dataset Directory Structure
- Save dataset as `./datasets/${PROJECT_NAME}`
- Annotation json format: COCO (Refer to [coco-format](http://cocodataset.org/#format-data))
> `${PROJECT_NAME}` is from [`./configs/initial_settings.json`](./configs/initial_settings.json)
```
# ./datasets/
${PROJECT_NAME}
│
│
├── labels
│ ├── train.json
│ └── val.json
│
│
└── imgs
│
├── train
│ │
│ ├── 0001.png
│ ├── 0002.png
│ ├── 0003.png
│ ├── ...
│
│
└──── val
│
├── 0001.png
├── 0002.png
├── 0003.png
├── ...
```
## COCO Evaluation
### Evaluation Script
```bash
python eval_coco.py --ckpt=${CKPT_PATH} --img_prefix=${IMG_PREFIX} --coco_gt=${COCO_GT_PATH}
```
#### Options
- `--ckpt`: Checkpoint file path
- `--img_prefix`: Image directory path to evaluate
- `--coco_gt`: COCO GT file path
- `--conf_thr`: Inference confidence threshold (Default: 0.05)
- `--img_exts`: Extensions of the image to evaluate. (Default=`['.png', '.jpg', '.jpeg']`)
If you put another image extensions, separate elements by comma (`--img_exts=.jpeg,.PNG`)
## Reference
### Paper
- YOLOv3: An Incremental Improvement ([arXiv Link](https://arxiv.org/abs/1804.02767))
```
@misc{redmon2018yolov3,
title={YOLOv3: An Incremental Improvement},
author={Joseph Redmon and Ali Farhadi},
year={2018},
eprint={1804.02767},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
### GitHub Repository
- [pythonlessons / TensorFlow-2.x-YOLOv3](https://github.com/pythonlessons/TensorFlow-2.x-YOLOv3)