Batch size and timing for single GPU setup about aa-rmvsnet

[QT-Zhu]

We set the batchsize of inferene at scripts/eval_dtu.sh. If you set this in env.sh, I doubt it would be overwritten by 2 since env.sh is executed before.
For inference, 170 seconds are longer than our experiments, but considering 1060 is a relative old model, it makes sense to me, as long as the predicted depth map is correct.

[KevinCain]

Thanks for your response. To answer you, the depth maps and fusion steps work fine based on the resulting .ply mesh.

Could you provide a few reference timings? I'd like to explore your technique but the timing I see is is orders of magnitude longer than MVS-Net and variants which unfortunately makes testing impractical.

[QT-Zhu]

For DTU evaluation, the inference time for one image is around 40 seconds. Since you can increase the batchsize, the averaged inference time could be reduced to 20+ seconds. Considering the ideology that RNN regularization trades time for space, time is indeed a bottleneck.

[KevinCain]

Thank you @QT-Zhu for relaying a general sense of your evaluation times, and your elegant point that time and space are once again entangled properties. ;-) On a practical level, I'm eager to do a few tests on a more capable system.

I also want to express my thanks for packaging your code in a succinct, easy to follow way. It is one of the few deep learning projects that compiled the first time!