Difficulties detecting long-thin objects

[TekayaNidham]

Thank you so much for your contributions.

I'm trying to train model to detect architectural objects and none of the models i train seems to detect long and thin object (as illustrated in the attached image).

• the dataset size is more than fair, it works fine on axis aligned object detection.
• Checked and manually retraced the annotation polygons and there no issue with that

  Already tried the following architectures :

• Oriented RCNN (ResNet 18 with full sized images, R(50, 101, 152 )with the splitting pre-processing and ResNeXt101)
• Rotated Faster RCNN (ResNet(50..152) and ResNeXt)
• Retinatnet (ResNet101)
• ROI Trans (R50 and R101)
• ReDet with R50

and none of these exceeded 0.4 mAP for that specific class.

Do you have any architecture recommendation or a change in the config i can do to improve my results?

[zytx121]

I guess you can add the anchor ratios from [0.5, 1.0, 2.0] to [1/7.0, 0.5, 1.0, 2.0, 7.0]. @yangxue0827 Do you have any good suggestion？

[TekayaNidham]

changing it to those specific ratios didn't help can you tell me how those ratios works so i can maybe try multiple combinations

[zytx121]

We will generate different ratios' anchors according to it. And you could refer to https://github.com/open-mmlab/mmdetection/blob/56e42e72cdf516bebb676e586f408b98f854d84c/mmdet/core/anchor/anchor_generator.py#L151