This repo releases the code and models of team "SIAT_MMLAB" for several large-scale scene recogntion challenges.
| Challenge | Rank | Performance |
|---|---|---|
| Places2 challenge 2015 | 2nd place | 0.1736 top5-error |
| Places2 challenge 2016 | 4th place | 0.1042 top5-error |
| LSUN challenge 2015 | 2nd place | 0.9030 top1-accuracy |
| LSUN challenge 2016 | 1st place | 0.9161 top1-accuracy |
Basically, we have made three efforts to exploit CNNs for large-scale scene recognition:
- We design a modular framework to capture multi-level visual information for scene understanding, called as MRCNN.
- We propose a knowledge disambiguation strategy to deal with the label ambiguity issue of scene recognition.
- We discover several good practices to train CNNs on existing datasets, like class balancing, hard sample mining.
The following report describes the technical detais:
Knowledge Guided Disambiguation for Large-Scale Scene Classification with Multi-Resolution CNNs
Limin Wang, Sheng Guo, Weilin Huang, Yuanjun Xiong, and Yu Qiao, in arXive 2016.
Models and code coming soon!