Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/pjh5672/YOLOv3
YOLOv3 implementation using PyTorch
https://github.com/pjh5672/YOLOv3
object-detection pytorch yolo yolov3 yolov3-pytorch
Last synced: 12 days ago
JSON representation
YOLOv3 implementation using PyTorch
- Host: GitHub
- URL: https://github.com/pjh5672/YOLOv3
- Owner: pjh5672
- License: gpl-3.0
- Created: 2022-11-23T15:10:56.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2024-02-26T22:01:11.000Z (4 months ago)
- Last Synced: 2024-02-28T22:59:08.008Z (4 months ago)
- Topics: object-detection, pytorch, yolo, yolov3, yolov3-pytorch
- Language: Python
- Homepage:
- Size: 2.06 MB
- Stars: 3
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Lists
- awesome-yolo-object-detection - pjh5672/YOLOv3
README
#
YOLOv3
---
## [Content]
1. [Description](#description)
2. [Usage](#usage)
2-1. [K-medoids Anchor Clustering](#k-medoids-anchor-clustering)
2-2. [Model Training](#model-training)
2-3. [Detection Evaluation](#detection-evaluation)
2-4. [Result Analysis](#result-analysis)
3. [Contact](#contact)
---
## [Description]
This is a repository for PyTorch implementation of YOLOv3 following the original paper (https://arxiv.org/abs/1804.02767).
- **Performance Table**
| Model | Dataset | Train | Valid | Size
(pixel) | mAP
(@0.5:0.95) | mAP
(@0.5) | Params
(M) | FLOPs
(B) |
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| YOLOv3
(Paper:page_with_curl:) | COCO | train2017 | val2017 | 416 | 31.0 | 55.3 | 61.95 | 65.86 |
| YOLOv3
(Our:star:) | COCO | train2017 | val2017 | 416 | 33.9 | 54.9 | 61.98 | 66.17 |
| YOLOv3
(Our:star:) | PASCAL-VOC | trainval2007+2012 | test2007 | 416 | 43.2 | 80.0 | 61.63 | 65.74 |
- **Pretrained Model Weights Download**
- [DarkNet-53 Backbone](https://drive.google.com/file/d/1duXcafb2QgORHDO1w-7E1UusLfWInwgA/view?usp=drive_link)
- [DarkNet-53-tiny Backbone](https://drive.google.com/file/d/13X39tcmNnYghvdBiosma7gsfwCxjsFQx/view?usp=drive_link)
## [Usage]
#### K-medoids Anchor Clustering
- You extract anchor box priors from all instances' boxes at first.
```python
python kmedoids_anchor.py --exp my_test --data voc.yaml --n-cluster 9
```
```log
2022-11-28 13:00:53 | Avg IOU: 68.97%
2022-11-28 13:00:53 | Boxes:
[[0.646 0.456 ]
[0.214 0.5135135 ]
[0.052 0.08108108]
[0.126 0.30133334]
[0.836 0.8613333 ]
[0.396 0.7147147 ]
[0.148 0.13333334]
[0.314 0.28266665]
[0.066 0.184 ]]
2022-11-28 13:00:53 | Ratios: [0.36, 0.42, 0.42, 0.55, 0.64, 0.97, 1.11, 1.11, 1.42]
```
#### Model Training
- You can train your own YOLOv3 model using Darknet-53 with anchor box from above step. If you wanna train YOLOv3 on your dataset from the scratch, add "--scratch" in training command like below.
```python
python train.py --exp my_test --data voc.yaml --multiscale(optional) --scratch(optional)
```
#### Detection Evaluation
- You can compute detection metric via mean Average Precision(mAP) with IoU of 0.5, 0.75, 0.5:0.95. I follow the evaluation code with the reference on https://github.com/rafaelpadilla/Object-Detection-Metrics.
```python
python val.py --exp my_test --data voc.yaml --ckpt-name best.pt
```
#### Result Analysis
- After training is done, you will get the results shown below.
```log
2022-11-26 22:11:13 | YOLOv3 Architecture Info - Params(M): 61.65, FLOPS(B): 65.74
2022-11-26 22:14:32 | [Train-Epoch:001] multipart: 442.6813 obj: 0.5345 noobj: 421.2506 txty: 0.3686 twth: 1.1349 cls: 17.2549
2022-11-26 22:18:47 | [Train-Epoch:002] multipart: 8.5325 obj: 0.7439 noobj: 0.1363 txty: 0.3173 twth: 0.5397 cls: 3.8199
2022-11-26 22:24:22 | [Train-Epoch:003] multipart: 7.4975 obj: 0.7797 noobj: 0.2258 txty: 0.3067 twth: 0.4024 cls: 2.6644
2022-11-26 22:30:01 | [Train-Epoch:004] multipart: 7.1354 obj: 0.7838 noobj: 0.2711 txty: 0.3000 twth: 0.3534 cls: 2.2920
2022-11-26 22:35:32 | [Train-Epoch:005] multipart: 6.8276 obj: 0.7748 noobj: 0.3078 txty: 0.2902 twth: 0.3096 cls: 2.0461
2022-11-26 22:41:07 | [Train-Epoch:006] multipart: 6.6838 obj: 0.7763 noobj: 0.3230 txty: 0.2851 twth: 0.2985 cls: 1.8956
2022-11-26 22:46:43 | [Train-Epoch:007] multipart: 6.5033 obj: 0.7582 noobj: 0.3741 txty: 0.2809 twth: 0.2736 cls: 1.7837
2022-11-26 22:52:17 | [Train-Epoch:008] multipart: 6.4126 obj: 0.7480 noobj: 0.3909 txty: 0.2741 twth: 0.2778 cls: 1.7296
2022-11-26 22:57:56 | [Train-Epoch:009] multipart: 6.2144 obj: 0.7327 noobj: 0.4325 txty: 0.2676 twth: 0.2619 cls: 1.5891
2022-11-26 23:03:32 | [Train-Epoch:010] multipart: 6.0665 obj: 0.7071 noobj: 0.4835 txty: 0.2646 twth: 0.2449 cls: 1.5379
2022-11-26 23:04:55 |
- Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.232
- Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.476
- Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.187
- Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
- Average Precision (AP) @[ IoU=0.50 | area= small | maxDets=100 ] = 0.001
- Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.038
- Average Precision (AP) @[ IoU=0.50 | area=medium | maxDets=100 ] = 0.103
- Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.287
- Average Precision (AP) @[ IoU=0.50 | area= large | maxDets=100 ] = 0.590
...
2022-11-28 12:44:51 | [Train-Epoch:249] multipart: 2.5484 obj: 0.2838 noobj: 0.7613 txty: 0.1396 twth: 0.0656 cls: 0.1629
2022-11-28 12:44:52 | [Best mAP at 220]
- Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.549
- Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.795
- Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.604
- Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.159
- Average Precision (AP) @[ IoU=0.50 | area= small | maxDets=100 ] = 0.314
- Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.331
- Average Precision (AP) @[ IoU=0.50 | area=medium | maxDets=100 ] = 0.563
- Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.598
- Average Precision (AP) @[ IoU=0.50 | area= large | maxDets=100 ] = 0.822
```
---
## [Contact]
- Author: Jiho Park
- Email: [email protected]