AerialDetection
AerialDetection copied to clipboard
Results Problem?
Hello, Really amazing work. While doing inference, i faced the results are not as good as shown. I tested using the pre trained models:
-
faster_rcnn_RoITrans_r50_fpn_1x_dota
-
faster_rcnn_RoITrans_r50_fpn_1x_dota1_5
Can You help Explain why?
What is the image source that you use? Have you conducted the multi-scale inference?
@dingjiansw101 i am using demo_large_image.py for inference. I have downloded these images from https://unsplash.com/s/photos/military-base-aerial-view
First, model trained with DOTA-v1.5 should be more accurate than the DOTA-v1.0. Second, I think there is a domain shift between your images and DOTA images, which may decrease the performance.