excuse me, can you provide the best model of vcoco? Thanks! about spatially-conditioned-graphs HOT 5 CLOSED

Hi @whyang78,

We didn't attach the V-COCO model because it takes very little time to train it. Based on the instructions, you should be able to reproduce the results fairly easily. Were you able to do so?

Fred.

from spatially-conditioned-graphs.

i set bs=4 and 4 gpus , learning rate=0.00005,i find the map of vcoco is 50.3. can you solve this problem， thank you

from spatially-conditioned-graphs.

hi，thank you for your answers！because my computer is not able to use batchsize=32，so it is difficult to reproduce the reaults.or can you provide the vcoco cache results of best model? thank you very much !

from spatially-conditioned-graphs.

Hi @whyang78,

The batch size is not a big problem. Your setup should be fine. One thing to point out is that, a few of the V-COCO action classes do not correspond to HOIs. Based on the number you posted, it is likely that you did not exclude those classes.

The image above is the results from InteractNet. Classes included in the red box are the ones you should average over. Make sure you do the same.

If I remember this correctly, the V-COCO script computes AP for a redundant class point. The AP for that class should be zero since it is never predicted. The number you get is the mAP for 25 classes with one of them being 0. So the correct performance should be 50.3 * 25 / 24 = 52.4. The more rigorous way to compute this is to sum up the APs for the 24 classes and compute the mean.

Let me know how it goes.

Fred.

from spatially-conditioned-graphs.

Hi @whyang78,

The batch size is not a big problem. Your setup should be fine. One thing to point out is that, a few of the V-COCO action classes do not correspond to HOIs. Based on the number you posted, it is likely that you did not exclude those classes.

The image above is the results from InteractNet. Classes included in the red box are the ones you should average over. Make sure you do the same.

If I remember this correctly, the V-COCO script computes AP for a redundant class point. The AP for that class should be zero since it is never predicted. The number you get is the mAP for 25 classes with one of them being 0. So the correct performance should be 50.3 * 25 / 24 = 52.4. The more rigorous way to compute this is to sum up the APs for the 24 classes and compute the mean.

Let me know how it goes.

Fred.

thank you very much! I will try again!

from spatially-conditioned-graphs.