Introduce a good independent resizing benchmark

rom1504 / img2dataset

Easily turn large sets of image urls to an image dataset. Can download, resize and package 100M urls in 20h on one machine.

MIT License

3.6k stars 333 forks source link

Introduce a good independent resizing benchmark #127

Open rom1504 opened 2 years ago

rom1504 commented 2 years ago

100k representative images (10 shards)
A table in .md with current numbers

Then use it to benchmark several solutions. For example:

Albumentation batching support

rom1504 commented 2 years ago

Current speed is around 80 image/s per i7 core

rom1504 commented 2 years ago

I did some benchmarks with the current code, and it seems cv2.imdecode is enough to use most of the currently used cpu

ikmckenz commented 2 years ago

Something to look into: Pillow-SIMD (and even plain Pillow) claim better performance than OpenCV (controlled through Python bindings) for most resizing tasks.

https://python-pillow.org/pillow-perf/