How to get IOP score for each proposal? about learn-to-cluster HOT 7 CLOSED

@eezxuan This repo currently only contains the IoU prediction, as it has not included GCN-S yet.
To get the IoP prediction, you can either (1) replace IoU with IoP or (2) train with both IoU and IoP. Both solutions require small modification on the code.

Specifically, for solution (1), when preparing the ground-truth for each proposal, we only need to modify this line to compute IoP instead of IoU.
For solution (2), in addition to the ground-truth preparation, we need to modify the network to predict IoU and IoP with two fc layers respectively.

from learn-to-cluster.

@eezxuan We have added the IoP prediction. To decouple IoU and IoP for better illustration, we implement solution (1) as discussed above. Feel free to have a try. :-)

from learn-to-cluster.

from learn-to-cluster.

Hello @yl-1993
Thanks for your great work and project.
Currently I also try solution (1) to train the GCN-D for getting the IoP score and I find a small problem. In train.py, since the build_dataset function only gets cfg.train_data as input, the det_label attr is always missing, and the default value 'iou' is used. In this way, setting det_label to 'iop' in train_cluster_det_iop_ms1m.sh doesn't change the training target to IoP score and the target is still the IoU score. And to solve this problem, I add the following line
setattr(cfg.train_data, 'det_label', cfg.det_label)
before sending cfg.train_data into build_dataset function, which works for me.

from learn-to-cluster.

@Dyfine Good catch! Do you want to create a PR for this?

from learn-to-cluster.

Hi @yl-1993
I just create a PR for this modification. I am sorry that I am not very familiar with the PR steps and please inform me if i did something wrong.

from learn-to-cluster.

@Dyfine Thanks. It looks good to me and I have merged it into the master branch.

from learn-to-cluster.