Deep neural networks (DNNs) have been widely used for closed-set recognition. In other words, they only recognize objects that have been seen in training. Can DNN be used in open-set recognition to identify database objects and reject novel unseen objects as unknown? if yes, how?
Dear Wasseem Al-Obaydy Wasseem Al-Obaydy,
Look the link, maybe useful.
https://arxiv.org/pdf/1802.04365.pdf
Conference Paper A bounded neural network for open set recognition
Regards, Shafagat
Dear Wasseem,
Look at the following papers:
Baktashmotlagh, M., Faraki, M., Drummond, T., & Salzmann, M. (2018). Learning Factorized Representations for Open-set Domain Adaptation. arXiv preprint arXiv:1805.12277.
Busto, P. P., Iqbal, A., & Gall, J. (2018). Open Set Domain Adaptation for Image and Action Recognition. IEEE transactions on pattern analysis and machine intelligence.
Yang, Y., Hou, C., Lang, Y., Guan, D., Huang, D., & Xu, J. (2019). Open-set human activity recognition based on micro-Doppler signatures. Pattern Recognition, 85, 60-69.
Schmidt, S., & Heyns, P. S. (2019). An open set recognition methodology utilising discrepancy analysis for gear diagnostics under varying operating conditions. Mechanical Systems and Signal Processing, 119, 1-22.
Hi Wasseem N. Ibrahem Al-Obaydy ,
I may start illustration from a famous picture. (I have attached it.)
Open-set recognition is different with the close-set setting. For open-set recognition, the key is to find a good and robust representation. The representation can transfer to the novel objects. For example, some early work usually use the attributes (i.e., has tail or not, has wings or not) to conduct open-set recognition.
For DNN, we usually extract the feature before the classification layer. (For example, we usually use the `pool5` activation of ResNet50.) We use this feature as the object representation. Because it has contained the important semantic clues for classification, which could easily transfer to the novel objects.
The application of the open-set recognition could be image retrieval (based on the feature ranking), person re-identification, face verification (based on the feature similarity of new faces). There are some related papers.
[1] Radenović, Filip, Giorgos Tolias, and Ondřej Chum. "CNN image retrieval learns from BoW: Unsupervised fine-tuning with hard examples." In European conference on computer vision, pp. 3-20. Springer, Cham, 2016.
[2] Zheng, Zhedong, Liang Zheng, and Yi Yang. "A discriminatively learned cnn embedding for person reidentification." ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) 14, no. 1 (2017): 13.
[3] Oh Song, Hyun, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. "Deep metric learning via lifted structured feature embedding." In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4004-4012. 2016.
We have lately published a paper in NIPS doing exactly that, using a simple trick to reduce the magnitude of the deep feature representation for unknown/unseen objects. You can have a look at https://papers.nips.cc/paper/8129-reducing-network-agnostophobia.pdf
- Badges
- Science topic