search

VITA-Group / UAV-NDFT

[ICCV 2019] "Delving into Robust Object Detection from Unmanned Aerial Vehicles: A Deep Nuisance Disentanglement Approach"
91 stars 33 forks source link

Bird view gives mAP=0 #6

Open Ben93kie opened 4 years ago

Ben93kie commented 4 years ago

Thank you for providing your implementations. After getting this repo to run a few error occured. Both the pretrained baseline model faster_rcnn_1_10_3960.pth and a model I trained on UAVDT yields mAP=0 for bird view. The overall mAP also is wrong (for almost all the categories). Could you shortly tell me, how I have to set

self._angles, self._altitudes, self._weathers

and

self._weather_to_ind, self._altitude_to_ind, self._angle_to_ind

as there are different version that are commented out. If I just take the uncommented rows it yields an key error. (f.ex. should I choose them as so: self._angle_to_ind = {'front-side-view': 0, 'front-view': 1, 'side-view': 0, 'bird-view': 2} ?)

Furthermore, you said that you were gonna "discard the foggy class". Does that mean you don't train and test on images labeled as foggy? If I include the foggy class and assign it to "day", the mAP is considerably worse. Thank you!

wuzhenyusjtu commented 4 years ago

Hi Ben, I think your evaluation is wrong. Before you quantitively evaluate the model, you can run a qualitative evaluation on the testing images by visualizing the bounding boxes. Also, we didn't discard the foggy videos. We relabel them as day or night.

Ben93kie commented 4 years ago

Thank you very much for your reply. I managed to get it working. Just a few follow-up question if you don't mind: -did you use the original image size (i.e. 1024 x 520) or did you use the configuration large_scale (effectively upscaling the image to aprox. 1300x800)? And is there a change in configuration for training and testing, respectively? -did you use any data augmentation (such as horizontally flipping)?

Thank you for your help!