Object Detection Progress Tracker
Last few years have seen great progress made in the domain of Object Detection. With powerful classification networks as backbone, more and more well-designed detection heads have been proposed to handle the dilemma of recognition and localization. This repo serves as a tracker of these progress.
2-stage
| Model | Backbone | Tricks | VOC | COCO | FPS | Paper | Date | Note | Code |
|---|---|---|---|---|---|---|---|---|---|
| R-CNN | VGG16 | - | 58.5 | - | - | Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation | - | CVPR2014 Oral | MATLAB |
| Fast R-CNN | VGG16 | - | 70.0 | 19.7 | - | Fast R-CNN | 15.04 | CVPR2014 Oral | caffe |
| Faster R-CNN | VGG16 | - | 73.2 | - | - | Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks | 15.06 | NIPS2015 | MATLAB |
| Faster R-CNN+++ | ResNet-101-C4 | 1,2,3 | 73.8 | 34.9 | - | Deep Residual Learning for Image Recognition | 15.12 | CVPR2016 Best Paper | |
| R-FCN | ResNet-101 | 4 | 80.5 | 29.9 | 6 | R-FCN: Object Detection via Region-based Fully Convoluational Networks | 16.05 | NIPS2016 | caffe |
| Deformable Conv | Aligned-Inception-ResNet | 1,3 | - | 37.5 | - | Deformable Convolutional Networks | 17.03 | ICCV2017 Oral | MXNet |
| Faster R-CNN w FPN | ResNet-101-FPN | - | - | 36.2 | - | Feature Pyramid Networks for Object Detection | 16.12 | CVPR2017 Poster | caffe2 |
| Faster R-CNN by G-RMI | Inception-ResNet-v2 | - | - | 34.7 | - | Speed/accuracy Trade-offs for Modern Convolutional object detectors | - | COCO206 winner | TesnsorFlow |
| Mask R-CNN | ResNeXt | 6 | - | 39.8 | - | Mask R-CNN | 17.03 | ICCV2017 Best Paper | caffe2 |
| Light-haed R-CNN | ResNet-101-FPN | 4,6 | - | 41.5 | - | Light-Head R-CNN: In Defense of Two-stage Object Detector | 17.11 | - | |
| Light-haed R-CNN | Xception | - | - | 30.7 | 102 | Light-Head R-CNN: In Defense of Two-stage Object Detector | 17.11 | - |
1-stage
| Model | Backbone | Tricks | VOC | COCO | FPS | Paper | Date | Note | Code |
|---|---|---|---|---|---|---|---|---|---|
| YOLO | Custom | - | 63.4 | - | 45 | YOLO: You Only Look Once | - | CVPR2016 Oral | darknet |
| SSD500 | VGG16 | - | 75.1 | 24.4 | 23 | SSD: Single Shot Detector | 15.12 | ECCV2016 Oral | caffe |
| SSD | ResNet-101 | - | - | 31.2 | 8 | DSSD: Deconvolutional Single Shot Detector | 17.01 | - | |
| YOLO v2 | DarkNet-19 | - | 78.6 | 21.6 | 40 | YOLO9000: Better, Faster, Stronger | - | CVPR2017 | darknet |
| DSSD | ResNet-101 | - | - | 33.2 | 6 | DSSD: Deconvolutional Single Shot Detector | 17.01 | - | |
| RON384++ | VGG16 | 3,7,8 | 77.6 | 27.4 | - | RON: Reverse Connection with Objectness Prior Networks for Object Detection | - | CVPR2017 | caffe |
| RetinaNet | ResNet-101-FPN | - | - | 39.1 | - | Focal Loss for Dense Object Detection | 17.08 | ICCV2017 Best student paper | caffe2 |
| RefineDet512 | VGG16 | - | 81.8 | 33.0 | 24.1 | Single-Shot Refinement Neural Network for Object Detection | 17.11 | - | caffe |
| RefineDet512+ | ResNet-101 | 3 | - | 41.8 | - | Single-Shot Refinement Neural Network for Object Detection | 17.11 | - | caffe |
Tricks list
- box refinement
- context
- multi-scale testing
- multi-scale training
- OHEM
- RoIAlign
- box voting
- flip
To be added
- FSSD
- DSOD
- GRP-DSOD
- [ ]
Disclaimer:
- Pascal VOC AP results are evaluated with IOU 0.5. Models are trained on VOC07+12, tested on VOC07.
- MS COCO results are reported with training on trainval35k, testing on test-dev set.
- All code repos are official implementation.
- Reported FPS are usually on GPU, not comparable actually.
My Notes(in Chinese) can be found here.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent