IDNet about person_search HOT 5 CLOSED

1. We use the cropped and resized (256x256) pedestrian images as inputs. related code
2. It is 5533
3. Yes, but the ID pretraining itself is much easier than the joint training process. As you can have ~256 different class samples in each minibatch, while in the joint training process, this number is typically less than 10.

from person_search.

@Cysu Thanks a lot for your prompt reply!
I'll try it out in this way - please correct me if something wrong:
step1. pretrain: finetune IDNet (with plain softmax layer) from ImageNet pretrained VGG16 model, batchsize ~256;
step2. train: finetune IDNet (with RSS layer) from pretrained model in step1, batchsize ~256;

from person_search.

Yes. You're right. Maybe the batch size could be tweaked. We used 128 in the original experiments.

from person_search.

Thank you for sharing. I have some questions about person_search. Does the class of background is a must? Can this adding the background class improve the re-id accuracy?

from person_search.

We hypothesize that having the background class can improve the performance, as the proposals will inevitably contain false positives. Will conduct some experiments to verify this. Thank you very much for the question.

from person_search.