As you see, we have provided a baseline which get 0.5375 mAP in the validation set.
The baseline is based on face recognition only, which means that we search persons, whose face can be detected, by face feature similarity and ignore the miss-detected faces.
If you want to re-implement the baseline, this repo. may be helpful: https://github.com/hqqasw/FaceTool .
Which dataset did you train your baseline face verification network? What's the size or scale of the dataset?
Posted by: yirong.mao @ June 12, 2018, 3:49 p.m.Which dataset did you train your baseline face verification network? What's the size or scale of the dataset?
Posted by: yirong.mao @ June 12, 2018, 3:49 p.m.Which dataset did you train your baseline face verification network? What's the size or scale of the dataset?
Posted by: yirong.mao @ June 12, 2018, 3:49 p.m.Very sorry for late reply.
The baseline is just an off-the-shell tool, which means that it does not use any data of this competition.
More specific, it is a Resnet-101 trained on Ms-Celeb-1M.
If you finetune this model on the training data of this task, you can easily beat the baseline.
By the way, the baseline only use "face" of the instances.
But actually there are many instances without frontal faces in the dataset.
So if you utilize other features beyond face, you can also easily beat the baseline.