Questions about TTA strategy

[hht1996ok]

Dear junjie, Thank you for your outstanding contribution. I used the GlobalRotScaleTranAll and CustomRandomFlip3D functions in MultiScaleFlip3D, pts_scale_ratio is set to [0.9, 1.0, 1.1] and flip&pcd_horizontal_flip&pcd_vertival_flip is set to True. However, the val result is worse than the base model, I do not know what is the problem with the setting Could you tell me what augments (rotation, scaling, flipping, etc.) you have used in TTA, and what is the nms threshold set.

[junjie18]

@hht1996ok You can check whether each single result is right first. In my experiments, all single result used in TTA should have similar performance.

[hht1996ok]

@junjie18 Thank you for your reply, In my experiments, the same increase was achieved when scale and flip were used alone, while the results decreased when they were used together.

[junjie18]

@hht1996ok nms is not a good choice in TTA, weighted box fusion is more useful here. However, I do not suggest you pay much attention on TTA or model ensemble. It's not an interesting, nor useful technique in my mind.

[hht1996ok]

@junjie18 Thank you for your advice. I agree with you.