Mmdetection: Question about multi scale testing

[thangvubk]

Thank you for sharing an awesome repo. I have a question relating to multi-scale testing.
If i am not wrong, we can perform multi-scale testing by changingimg_scale=(1333, 800) in test dict.
However when i change img_scale to img_scale=[(400, 1400), (1200, 1400). I got worse mAP compare to single scale testing.

Am i wrong somewhere? Can you show me the correct way to perform multiscale testing in your repo?
The below is the results that i tried:
Model: Faster-R-50-FPN-1x

• Single-scale testing: 36.4 mAP
• Multi-scale testing: 35.5 mAP

Thank you!

[hellock]

The choice of multiple scales also matters. The two scales you choose is either two small or two large. If you choose [(1400, 1000), (1400, 600)], the results should be much better. Of course you can add more scales for better performance.

[thangvubk]

Thank you for the quick reply. If i choose [(1400, 1000), (1400, 600)]. The test scale will be randomly sampled between (1400, 1000) and (1400, 600) or it just is counted as two fixed scales.

[hellock]

During inference, scales are not randomly sampled, just the fixed values you specified.

[thangvubk]

I totally understand now. Thank you!

[chenxu93]

@hellock it seems this code base only support 2 scales when training and testing? It is right, and what are the differences between range model and value model?

[hellock]

The value mode supports arbitrary scales, and the range mode only needs two scales (min and max). The code should be already very clear.

[01lin]

@thangvubk I have a question about multi-scale testing. how to set multi-scale-testing?
when i change img_scale=[(1000,600),(834,500)], i got the wrong code: TypeError: Scale must be a number or tuple of int, but got .
Can you help me ? Thank you~

[01lin]

@thangvubk I have a question about multi-scale testing. how to set multi-scale-testing?
when i change img_scale=[(1000,600),(834,500)], i got the wrong code: TypeError: Scale must be a number or tuple of int, but got .
Can you help me ? Thank you~

I used the official test-image script:

[chenxu93]

@01lin if you use the official test-image script: it only support one scale when you perform inference. If you want to perform multi-scale test, you should use test command like:
python tools/test.py --gpus --out