Can we change batch-size and num-iters in test-time optimization? about humor HOT 2 CLOSED

The batch size you can use depends on your GPU. Using iMapper sequences that are 2 sec (60 frames) long (hence --imapper-seq-len 60 in the config), a batch size of 32 will jointly optimize over 1 minute of video which is probably too much to fit on memory. Changing from --num-iters 30 80 70 to --num-iters 3 8 7 for these 2 sec sequences will give very bad results because the optimization will not have enough iterations to converge.

Alternatively, you could make the sequence lengths much smaller, like --imapper-seq-len 10 with a larger batch size and a smaller num-iters (since convergence is faster for short sequences). However this is not recommended because the method will likely fail for small sequences where the person is occluded - the first pose of the motion sequence will be initialized from a frame with heavy occlusions which usually severely hurts performance. You'll have to experiment and see the results.

Note that any time you change --num-iters or seq-len, you will also need to change all the --stage3-tune-init-* arguments. For example, if --num-iters 30 80 70, then --stage3-tune-init-freeze-start and --stage3-tune-init-freeze-end must be < 70 since these define the different steps performed in the third stage of optimization (this is detailed in the supplementary material of the paper). And --stage3-tune-init-num-frames 15 must be < seq-len.

Unfortunately, besides using a huge GPU and increasing batch size I don't think you will be able to improve optimization time too much.

from humor.

Closing for now. Let me know if you have any other issues.

from humor.