Open Tyhye opened 7 years ago
Basically, you can already detect "smaller object" using the current architecture and scales but only with low confidence. If you lower the threshold, smaller objects can also be considered as detections.
The main problem could be there's not enough "small objects" in the dataset. If you have a try on the other dataset eg, Caltech Pedestrian, that will increase the performance on small object detection.
Hello.May I change the min_ratio (in ssd_pascal.py ) to train my own model and detect smaller object? My English is poor... thank you