This is the implementation of the "Joint Neural Networks for One-shot Object Recognition and Detection" thesis by Camilo Vargas.
- Python 3.x
- Numpy
- OpenCV
- Pytorch
- Matplotlib
- PIL
- QMUL-OpenLogo dataset (https://qmul-openlogo.github.io/)
Set the training and testing parameters in python params/config.py file. Run the python main.py file to train/test the defined configuration.
Joint neural network performance for one-shot object detection tested on the VOC dataset, leaving four unseen classes:
| Method / class | cow | sheep | cat | aeroplane | mAP |
|---|---|---|---|---|---|
| JNN | 64.7 | 51.0 | 65.2 | 43.5 | 69.1 |
Joint neural network performance for one-shot object detection trained on the COCO dataset, tested on the VOC dataset:
| Method / class | plant | sofa | tv | car | bottle | boat | chair | person | bus | train | horse | bike | dog | bird | mbike | table | cow | sheep | cat | aero | mAP |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| JNN | 9.5 | 69.3 | 49.8 | 60.3 | 7.2 | 29.1 | 10.1 | 6.7 | 60.3 | 57.2 | 58.5 | 45.3 | 62.6 | 45.6 | 74.8 | 29.0 | 70.4 | 55.4 | 54.4 | 88.1 | 47.1 |
Testing on the top performing Open-Logodataset classes and the mAP results for the whole dataset
| Class | AP / mAP |
|---|---|
| anz_text | 100.00 |
| rbc | 98.86 |
| blizzardentertainment | 98.34 |
| costco | 93.26 |
| 3m | 90.61 |
| bosch_text | 90.00 |
| gap | 89.47 |
| lexus | 88.92 |
| generalelectric | 83.32 |
| hp | 82.81 |
| levis | 79.36 |
| airhawk | 79.17 |
| danone | 79.02 |
| armitron | 77.73 |
| 77.66 | |
| all | 52.84 |
Joint Neural Networks for One-shot ObjectRecognition and Detection. Camilo Jose Vargas Cortes. School of Electronic Engineering and Computer Science. Queen Mary University of London. 2020.
This code is based on the following repositories:
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent