The model can only predict one class, but UA-DETRAC has 4 classes. How do you evaluate the AP?

hu64 / SpotNet

Repository for the paper SpotNet: Self-Attention Multi-Task Network for Object Detection

MIT License

51 stars 10 forks source link

The model can only predict one class, but UA-DETRAC has 4 classes. How do you evaluate the AP? #7

Closed sisrfeng closed 4 years ago

sisrfeng commented 4 years ago

Many thx!

hu64 commented 4 years ago

UA-DETRAC's ground truth has 4 classes annotated, indeed. However, the evaluation takes only one class (You can check in their matlab toolkit for more details). The AP is significantly higher if you train with only one class to evaluate on one class instead on training on four.