[Docs] The crop stride in paper and code

shaunyuan22 / SODA-mmdetection

SODA-D Small Object Detection Toolbox and Benchmark

https://shaunyuan22.github.io/SODA/

Apache License 2.0

33 stars 8 forks source link

[Docs] The crop stride in paper and code #10

Closed promefire closed 1 year ago

promefire commented 1 year ago

Branch

master branch https://mmdetection.readthedocs.io/en/latest/

📚 The doc issue

In the paper ,I see "Directly feeding the high-resolution images in SODA to deep model is infeasible due to the GPU memory limitation,hence we crop original images into a series of 800 × 800 patches with a stride of 650" But in tools/img_split the stride is 150 so ,which one should i choose

Suggest a potential alternative/fix

No response

shaunyuan22 commented 1 year ago

just run the script by default then you get the patches with a stride of 650. in /tools/img_split/sodad_split.py, the argument stepSz is NOT the stride which follows the equation: patch_size - stepSz = stride.