Incremental Learning in Person Re-Identification
- 0 Collaborators
Person Re-Identification is still a challenging task in Computer Vision due to variety of reasons. On the other side, Incremental Learning is still an issue since Deep Learning models tend to face the problem of overcatastrophic forgetting when trained on subsequent tasks. In this paper, we propose a model which can be used for multiple tasks in Person Re-Identification, provide state-of-the-art results on variety of tasks and still achieve considerable accuracy later on. We evaluated our model on three datasets Market 1501, CUHK-03, Duke MTMC. Extensive experiments show that this method can achieve Incremental Learning in Person ReID efficiently as well as for other tasks in computer vision as well. ...learn more
Project status: Published/In Market
Groups
Student Developers for AI
Intel Technologies
AI DevCloud / Xeon,
MKL,
Intel Python
Overview / Usage
Deep neural networks have revolutionized the field of
computer vision. In recent years, a lot of work has been
done in Person Re-Identification ,we’ve seen a considerable
progress but still we face a lot of challenges in terms of getting
accurate predictions in real life instances.It plays an important
role in many areas, surveillance being one of them.
In some sense, it can be compared to other prominent tasks
in computer vision like Image Retrieval or Object Detection,
where a lot of progress have been made. Moreover,
there has been a growing demand of deep learning models
that incur low computational cost. Deployment of such
models can be cumbersome and may not prove to be much
efficient especially if the same task can be carried out with
lesser number of parameters. Given a set of images of a person
taken from different angles from different camera, our
model is required to generate a higher prediction if those
images are of the same person and vice versa. The problem
is composed by multiple reasons some of which may
include background clutter, illumination conditions, occlusion,
body pose, orientation of cameras. Numerous methods
have been proposed to address some of these issues. So
far the models that have been proposed in Person ReID are
good in doing well in particular dataset and when tested on
quite dissimilar dataset, they struggle to get just right predictions.
Unlike other tasks such as Image Classification or
Object Detection, we are required to have our model perform
well on a large number of classes and all these images
are not as much distinctive as other objects do which makes
it difficult for neural net to predict. We devise a new method
that can be used to create robust Person-ReID systems at
lower computational cost that can not only perform well on
one tasks, but if trained properly using our techniques, can
be well adapted to other tasks as well.
A detailed overview of the some part of research work can be found on this Intel Dev Zone Blog ( https://software.intel.com/en-us/articles/part-1-using-transfer-learning-to-introduce-generalization-in-models)
Methodology / Approach
Discussed in paper (https://arxiv.org/pdf/1808.06281.pdf)
Technologies Used
Intel Optimized Python Distribution, Intel MKL, Intel Nervana DevCloud
Repository
https://github.com/prajjwal1/person-reid-incremental