issue about metrics and VG datasets

Maelic / SGG-Benchmark

A New Benchmark for Scene Graph Generation, targeting real-world applications

MIT License

34 stars 5 forks source link

issue about metrics and VG datasets #16

Closed focus1224 closed 2 weeks ago

focus1224 commented 3 months ago

Hi, I'm new to SGG. Thank you for sharing this incredibly useful project. I have a couple of questions: 1.Where can I find the code for calculating the evaluation metrics? 2.While using the VG dataset, I noticed that some images seem to be corrupted. For example, images 2397804.jpg, 2347067.jpg, and 2344194.jpg. How did you handle these corrupted images? Could this be an issue with my download? Thank you!

Maelic commented 3 months ago

Hi @focus1224, the metrics are computed here: https://github.com/Maelic/SGG-Benchmark/blob/main/sgg_benchmark/data/datasets/evaluation/vg/sgg_eval.py

If you find more corrupted images for VG, you can add their names to this list: https://github.com/Maelic/SGG-Benchmark/blob/6b4c713c990d5e7ff2e2421b2e8a5e1142889a0f/sgg_benchmark/data/datasets/visual_genome.py#L422

Maelic commented 2 weeks ago

Closing due to inactivity