1.Basic model selection based on the SIFT features:
Using the SIFT features provided by the professor, we have try SVM-Kernel, Random Forest, GBM and KNN four classification models.
The error rate result:
SVM: 43% (Some parameters need to be improved later and we guess that the error rate will reach about 30% in the end)
KNN: 32%
Random Forest: 26%
GBM: 26%
So, based on the above results, we think that improving the parameters of model or the selection of model may not have too much impact on the outcomes.
Currently, we decided to use GBM or simple combinations of several models.
2.Features related problems
As we all known, the improvement of features is quite important. But we do not have the detailed methods and ideas now.
We decided to try google some new or improved methods of feature abstraction except the OpenCV package mentioned by the professor.
Currently, this part is the most significant one in our project.
Jobs assigned
Liangbin to improve the selected models and run the abstracting new features on the trained model
Yaqin, Zheren and Tongyue to improve features abstraction methods and get the feature data before Wednesday noon.
And Imroze you can choose whether to improve the model part or the feature part.
The above part are supposed to be done before our next group meeting at Uris, Wednesday 3:00 pm.
All group members will working on the final report together on Thursday.
is2548
commented
7 years ago
1.Basic model selection based on the SIFT features: Using the SIFT features provided by the professor, we have try SVM-Kernel, Random Forest, GBM and KNN four classification models. The error rate result: SVM: 43% (Some parameters need to be improved later and we guess that the error rate will reach about 30% in the end) KNN: 32% Random Forest: 26% GBM: 26% So, based on the above results, we think that improving the parameters of model or the selection of model may not have too much impact on the outcomes. Currently, we decided to use GBM or simple combinations of several models.
2.Features related problems As we all known, the improvement of features is quite important. But we do not have the detailed methods and ideas now. We decided to try google some new or improved methods of feature abstraction except the OpenCV package mentioned by the professor. Currently, this part is the most significant one in our project.