inference.py, eval.py, demo.py all buggy!

[vinjohn]

Hi, the inference.py, eval.py and demo.py are all buggy in this repo, would you please update a version which can really work?
The train.py is fine.

[yangxue0827]

can you give me the error information. @vinjohn

[vinjohn]

for R2CNN, everything is fine.
for RRPN, take inference.py for instance:

the build_whole_detection_network need 3 input and return 6 output.

det_net.build_whole_detection_network(input_img_batch=img_batch, gtboxes_h_batch=None, gtboxes_r_batch=None)

however, the implementation of build_whole_detection_network in libs/networks/build_whole_network.py:

build_whole_detection_network(self, input_img_batch, gtboxes_batch)
2 input and 4 return (final_boxes, final_scores, final_category, loss_dict)

the exact error message I get is:
TypeError: build_whole_detection_network() got an unexpected keyword argument 'gtboxes_h_batch'

[yangxue0827]

please update your code and try again. @vinjohn

[vinjohn]

By the way, I have another question, please help.

The default config for DOTA with base anchor = 256, anchor scales = [0.0625, ..., 2.0], anchor ratios = [1, ... ,1/7]
I think this setting is meant for small and long objects in DOTA.
However, in my own dataset, the default anchor size in faster rcnn is fine (128, 256, 512). I know how to change anchor ratios list back to normal, but not sure about base_anchor_size_list, anchor_scales and roi_scale_factors.

Could you please give me some instructions on config these three parameters?

[vinjohn]

This time the inference is going well, but I think a nms process should be added, anyway, I add it myself.

The anchor scale question, please help, thanks a million!

[yangxue0827]

base anchor * anchor scales = [16, 32, 64, 128, 256, 512] as the base_anchor_size_list. Anchor ratios = [1, ... ,1/7] controls the shape of the anchor under the same base. roi_scale_factors controls the value of predict offset in a small range. @vinjohn