Out-of-bounds coordinates in annotation files P0334.txt and P1872.txt

[davidgeorgewilliams]

Description

While processing the DOTA 1.5 dataset, we've encountered annotations with coordinates that lie outside the boundaries of their corresponding images. This issue has been observed in at least two files: P0334.txt and P1872.txt.

Details

• File: P1872.txt Image dimensions: 2836×1689 Example problematic annotation:

235.0 1768.0 253.0 1765.0 259.0 1809.0 241.0 1810.0 small-vehicle 0

• File: P1872.txt Image dimensions: 4033x4096 Example problematic annotations:

  10866.0 4726.0 10822.0 4862.0 10646.0 4822.0 10700.0 4666.0 plane 0 11012.0 4952.0 11056.0 4824.0 11192.0 4876.0 11162.0 4994.0 plane 0 11084.0 4796.0 11132.0 4652.0 11298.0 4714.0 11252.0 4850.0 plane 0 9992.0 1322.0 9922.0 1116.0 10120.0 1056.0 10190.0 1258.0 plane 0 10076.0 1500.0 9920.0 1558.0 9864.0 1414.0 10030.0 1358.0 plane 0 10256.0 1426.0 10096.0 1480.0 10052.0 1322.0 10206.0 1270.0 plane 0 10506.0 1334.0 10346.0 1396.0 10282.0 1236.0 10446.0 1174.0 plane 0 10738.0 1226.0 10606.0 1272.0 10556.0 1130.0 10682.0 1080.0 plane 0 11300.0 1084.0 11396.0 1056.0 11430.0 1162.0 11344.0 1204.0 plane 0 11376.0 988.0 11278.0 1014.0 11224.0 896.0 11324.0 864.0 plane 0 11500.0 932.0 11408.0 962.0 11362.0 852.0 11466.0 816.0 plane 0 11568.0 1036.0 11658.0 1000.0 11688.0 1078.0 11602.0 1114.0 plane 0 11590.0 260.0 11734.0 204.0 11782.0 366.0 11652.0 418.0 plane 0 11178.0 404.0 11332.0 354.0 11410.0 526.0 11254.0 590.0 plane 0 11036.0 496.0 11112.0 462.0 11154.0 534.0 11076.0 574.0 plane 0

Questions

1. Is this a known issue in the DOTA 1.5 dataset?
2. How does this repository recommend handling out-of-bounds coordinates?
3. Are there any plans to address or correct these annotations in future releases?

Impact

These out-of-bounds coordinates can lead to issues when converting annotations to other formats (e.g., YOLO) or when training object detection models. They may result in incorrect object positions or sizes, potentially affecting model performance.

Proposed Solutions

We're considering the following approaches to handle this issue:

1. Clipping coordinates to image boundaries
2. Scaling coordinates to fit within image dimensions
3. Filtering out annotations with out-of-bounds coordinates

Any guidance on the preferred method would be appreciated.

Additional Information

• DOTA version: 1.5
• Processing script: (If applicable, provide a link or description of your processing script)

Thank you for your attention to this matter. We'd appreciate any insights or recommendations on how to proceed.