Thoughts about deeplearning
Deep learning is quite popular in computer science. Some think deep learning model is a black box. However, I think some part of deel learning model can be explicitly explained.
What is deep learning?
Compared with conventional machine learning, deep learning models have much more complex network architecture and act more like human. Why am I say that the deep learning models are anthropoid? Deep learning models are unlike conventional machine learning, which is fed by handcraft features and ap ...
Locally Aware Transformer
Review of Person Re-Identification with a Locally Aware Transformer
Summary in one sentence.
Sharma et al. propused a locally aware transormer inspired by Parts-based Convolution Baseline(PCB) and ViT(Vision Transformer) as well as a fine-tunning strategy which further improve Re-ID accuracy
Introduction
Person Re-Identification task:
Person Re-ID task includes object dectation and image retrival tasks in earily researches. It separates the tasks and focuses on finding the most similar images i ...
Summary of approaches used in person Re-Id task.(part 1)
Deep ReID Architecture Types
Classification Models
Classification models consider ReID as a multi-class classification problem. Those models use the softmax loss to predict the class of an input query. Softmax loss encourages the separation of different classes but struggles with large intra-class variations.
Several methods overcome the inability of softmax loss to handle intra-class variations by introducing new loss.
In “A loss combination based deep model for person re-identification”, Zhu ...