Hello, I have the following questions, looking forward to your answers. about sigma HOT 3 CLOSED

Hi, thanks for your interest！

Q1. Yes, the available results should be in the range from 42 to 44. The reported results are not the best ones in our experiments. For example, we have 45AP50 model for C2F with two–stage training. We use more validation only to ensure the easy reproductability for different experimenal environments in an end-to-end manner.
Q2. Actually, this will be similar for AP50 but not for the more strict metrics, e.g. AP. Training more iterations mainly improves the overall performance of the object detector.
Q3. For your own datasets, I think both settings are reasonable since some DAOD benchmarks have an extra test split.

from sigma.

@wymanCV
Thank you very much for your reply. I would like to ask if you have ever encountered that sometimes the maximum accuracy of AP50 saved is obviously 4 points lower than the previous accuracy when it is reproduced, and every 2500 iterations is also about 4 or 5 points lower than the previous iteration? What's going on here? Thank you. Sometimes this is normal when code is reproduced.

from sigma.

Hi, I haven't encountered such significant instability after training enough iterations. But I think it is normal for adaptative FRCNN in the early training stage, as shown here.

from sigma.