Anchor boxes have been the prevalent way to generate candidates for the ground truth bounding boxes in the object detection problem. Yet, this approach is such a hassle and downright confusing. This tutorial leverages an object detection method named FastestDet that is lightweight and anchor-free. PASCAL VOC 2007 and 2012 datasets are utilized to evaluate the model's capability. Here, the train and validation sets of PASCAL VOC 2012 are used for the train and validation while the test set of PASCAL VOC 2007 is allotted for the testing phase in this tutorial. Eventually, the inference set (the test set of PASCAL VOC 2007) is used to see the qualitative performance of the model.
Explore here to execute training, testing, and inference.
The table below presents the quantitative result of the model on the test set.
| Test metric | PASCAL VOC 2007 |
|---|---|
| Loss | 3.058 |
| mAP@0.5:0.95 | 15.307% |
Loss curve on the train set and the validation set.
The qualitative results of the model on the inference set are shown below.
Two motorbikes (left), a person and a horse (middle), and a car and an aeroplane (right) are detected.