| 研究課題/領域番号 |
15K16024
|
|---|
| 研究機関 | 奈良先端科学技術大学院大学 |
|---|
研究代表者 |
伍 洋 奈良先端科学技術大学院大学, 研究推進機構, 助教 (30750559)
|
|---|
| 研究期間 (年度) |
2015-04-01 – 2018-03-31
|
| キーワード | Person re-identification / Set-based recognition / Deep learning / Metric learning / Efficiency / Scalable / Transfer learning |
|---|
| 研究実績の概要 |
We published 1 international journal paper and 2 international conference papers which are related to this project.
In the international journal paper entitled "Locality based discriminative measure for multiple-shot human re-identification", we proposed a new set-to-set dissimilarity which cares about both majorities and minorities of samples in the set pairs, and explored a local metric field to make the best use of such a dissimilarity for combining locality and metric learning for set-based re-identification. This work is an important component for the whole proposed model.
In the two international conference papers, we investigated hierarchical feature learning and its combination with deep features from convolutional neural networks. This is a new direction that we have explored for borrowing the latest progress from deep learning to enhance our proposed model. The two papers are just some preliminary studies.
Besides that, we have also submitted several other papers (3 journal papers and 1 conference paper) for reviewing, and we are still waiting for their acceptance. Their results will be reported next year.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
3: やや遅れている
理由
We have got encouraging research progress and achievements, and explored new possibilities that we didn't planned to do. However, we weren't able to finished all the planned research for the first year, because:
1. Since the beginning of this project, the principle investigator (PI) has been working at a new position (the actual research staff that takes care of the newly built NAIST International Collaborative Laboratory for Robotics Vision) for establishing the international collaboration and starting several new research topics. The new research topics need more efforts than expected in the beginning, so the work on the project is delayed a little bit.
2. Some new research trend on the research topic appeared and developed very quickly, so the PI spent some time investigating it. More concretely, deep learning models have recently shown striking performances on many recognition problems and also achieved significant better performance than other methods this year on the person re-identification tasks. Therefore, the PI was trying to catch up its latest progress and borrow some key ideas from it for enhancing our research. We also got some initial research outcomes from the study.
|
| 今後の研究の推進方策 |
Though the original plan has been delayed a little bit, we still believe that the proposed model is a promising solution. Meanwhile, the latest research progresses can also be adopted to enhance parts of our model. Therefore, the plan for our future work in the left two years will be as follows.
1. To test the state-of-the-art features, including the ones from deep neural networks.
2. To keep implementing the fast clustering of data and fast search techniques.
3. To try some recent metric learning models for a better integration with our collaborative representation model.
4. Doing the experiments not only on our own dataset, but also on a newly published dataset which contains significantly more people.
|
| 次年度使用額が生じた理由 |
For the first year, most of the traveling cost has been covered by other budgets of the university, so that more money will be reserved for supporting the publications and presentations of our research results at conferences venues for the next two years.
|
| 次年度使用額の使用計画 |
The reserved budget will be used for supporting probably more publications and travelings in the next two years. Meanwhile, we will use some budget for buying a few new PCs to support our research.
|
