• Stars
    star
    174
  • Rank 219,104 (Top 5 %)
  • Language
    MATLAB
  • License
    MIT License
  • Created almost 8 years ago
  • Updated almost 7 years ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

Implement for ``Learning Deep Features via Congenerous Cosine Loss for Person Recognition''

COCO Loss

Update: we have released the network structure for face recognition. And we answer some common questions in Q&A.

Update: we have an updated version of this work, which we call COCO_v2 [arXiv]. The difference from v1 to v2 is that, in a short story, we (a) propose an optimal scale factor imposed on the input feature before COCO loss; (b) experiments are extended to large-scale face recognition datasets beyond PIPA.

The codebase is UPGRADED to COCO_v2. :)

Codebase for Learning Deep Features via Congenerous Cosine Loss for Person Recognition (COCO_v1), [arXiv]. In the paper, we use the newly proposed PIPA dataset for evaluation; it is a large-scale dataset for person recognition.

We provide a sample code for generating distributions of COCO Loss, Softmax Loss and Center Loss as Figure 3 in our paper. Please follow the steps below:

Runing sample code

  1. Compile CaffeMex_v2 with matlab interface

  2. Link (Linux) or copy (Windows) the path CaffeMex_v2/matlab/+caffe to the root directory in this project

  3. Run script_show_minist.m

  4. It takes just few seconds to generate three figures as:

Notice that the distributions are generated based on caffemodels in model/[LOSS_NAME]. You can reproduce the result by using network definition files in model/train_prototxts/.

The proposed COCO loss layer (known as center_projection_layer.cu here) and the normalized layer (here) are two operations related wiht COCO. The gradient derivation in the initial release (v1) on arXiv is wrong; please resort to v2 (latest version): Eqn.(8) in v2 paper corresponds to the implementation around lines here.

Q&A

Here we list some commonly asked questions we received from the public. Thanks for your engagement to make our work matter!

  • Relationship of person recognition task with face recognition, person Re-ID.

    Face recognition is the traditional vision task to recognize a person's identify based solely on face (often with some setting constraints, like frontal faces); whereas the context of person recognition carries more. The main difference between person recognition and re-id resides in the data logistics. The former is to identify the same person across places and time. In most cases, the identity varies a lot in appearance under different occasions. The latter is to detect person in a consecutive video, meaning that the appearance and background do not vary much in terms of time.

  • Training curve looks weird. Any tricks to make COCO work?

    Hah, good question. There are mainly two. First is we initialize the network with a pre-trained model for the re-id task in order for the loss to converge well in a large-scale dataset like PIPA; second is we do not normalize the feature to 1. There are some numerical problems for gradients of centroids if we normalize all features and weights to 1. In MNIST, we initialize the scaler to 1 for weights and 2 for features while in PAPA we set 30 and 100, respectively.

  • [New] How could COCO achieves 99.86% on LFW with 10 incorrect annotations?

    Some people represent the issue of 'upper limit accuracy' of LFW such as this issue. We have to indicate that there are 10 incorrect pairs on LFW but only the 6 matching pairs are 'really incorrect'. The faces in the other 4 mismatching pairs are labeled with wrong identities but they are still mismatching. So the upper limit accuracy is (6000-6)/6000=99.90%. For more details please refer to Errata of LFW.

  • [New] Any comments on the evalutation of LFW and MegaFace?

    We believe that both LFW and MegaFace are not precise benchmark for face recognition/verification. For LFW, most of the recent CNN based algorithms get the saturation area of this benchmark. For MegaFace, the top-1 accuracy introduces an inverse correlation with the accuracy of face detector. That is, the worse of your face detector, the less real faces you will detect in distractors, so there will be more meaningless background in distractors and the top-1 accuracy will be higher. In our work, we use the commercial version of RSA to be our detector and we ensure that all the faces we detected are real faces. For the images we didn't detect a face, none of them includes faces larger than 10x10 (in fact, most of them are entire background). Above all, we appeal to the researchers in area of face recognition to pay more attention on ablation study but not the absolute value of accuracy.

  • [New] Will the models for face recognition be released in future?

    The network structure is released but the trained model will not be released at least in recent days. Since the alignment algorithm and the cleaned MS1M are classified and copyright by Sensetime Group Limited.

Still having questions?

Feel free to drop us an email sharing your ideas.

Related work

Please refer to the paper for details.

  • Feature Incay for Representation Regularization, [link], arXiv, 2017.
  • Pose-Aware Person Recognition, [link], CVPR, 2017.
  • Cosine Normalization: Using Cosine Similarity Instead of Dot Product in Neural Networks, [link], arXiv, 2017.
  • Large-Margin Softmax Loss for Convolutional Neural Networks, [link], ICML, 2016.

Citation

Please kindly cite our work in your publications if it helps your research:

@article{liu_2017_coco_v2,
  Author = {Liu, Yu and Li, Hongyang and Wang, Xiaogang},
  Title = {Rethinking Feature Discrimination and Polymerization for Large-scale Recognition},
      journal={arXiv preprint arXiv:1710.00870},
      Year = {2017}
}

@article{liu_2017_coco_v1,
  Author = {Liu, Yu and Li, Hongyang and Wang, Xiaogang},
  Title = {Learning Deep Features via Congenerous Cosine Loss for Person Recognition},
  Journal = {arXiv preprint: 1702.06890},
  Year = {2017}
}