blue-oil/blueoil

Coordinates of gt_boxes are invariant after augmentation related to coordinates transformation

Open

#60 建立於 2018年12月5日

在 GitHub 查看
 (0 留言) (2 反應) (0 負責人)Python (250 star) (85 fork)batch import
buggood first issuehelp wanted

描述

Coordinates of gt_boxes are invariant after augmentation related to coordinates transformation, "Pad", "Rotate" and "Crop". Here is the examples on "Pad". The coordinates of gt_boxes and shifted images are not match:

On "Pad", I fixed it by adding the code to __call__method like this:

        gt_boxes[:,0] = gt_boxes[:,0]+self.left
        gt_boxes[:,1] = gt_boxes[:,1]+self.top

and modifiiied the argument and the return value accordingly.

Then, it works well.

Likewise, I think "Rotate" and "Crop" should be modified accordingly.

貢獻者指南

技術棧
python
領域
machine learningdata
議題類型
bug
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
2
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
1-3 hours
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
stale
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
clear
前置要求
PythonNumPyObject detection augmentation
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
75
研究方向
The issue reports that after Pad, Rotate, and Crop augmentations, bounding box coordinates are not transformed. The reporter provided a fix for Pad by adding coordinate offsets. For Rotate and Crop, similar modifications are needed: adjust the gt boxes coordinates according to the transformation parameters. Examine the augmentation implementations in the BlueOil codebase (likely in data augmentation modules) to apply the correct coordinate transforms. Ensure that the fix is consistent across all three augmentations and test with sample images to verify alignment.