Now I am implementing bag of words method on 1 million image data-set. For getting 20000 words or class centers, I use 2 million sift features . How ever it is too slow to getting centers of each class.
If anyone have experience of using KNN on large data-set, I hope sharing your experience ...
I have used CUDA + KNN. Very powerful. You can also do feature selection with it. See https://www.researchgate.net/publication/262949628_Fast_Feature_Selection_in_a_GPU_Cluster_Using_the_Delta_Test?ev=prf_pub
Article Fast Feature Selection in a GPU Cluster Using the Delta Test
Hi, recommend you to use an approximate knn. In the naive knn implementation you have to compute the distance from all the training samples against all the testing samples. Using an aproximate implementation that works preaty well you can speed it up. The standard implementation is FLANN:
http://www.cs.ubc.ca/research/flann/
Good luck!!
data sets used in experimental sdudies:
Reference Corpus size
Hulth A. 800
Nguyen and Kan 120
Wan and Xiao 308
Schutz 500
Krapivin et al. 680
Su Nam Kim 100
I know that there exist a KNN implementation in CUDA. But, this implementation uses the naive KNN implementation not the fast approximation as in FLANN. You can find it here:
http://vincentfpgarcia.github.io/kNN-CUDA/
have you used the Flann, I have downloaded it, and implement it.. it gives you the nearest neighbor however i want it to make classification not nearest neighbor.. how can I implement it ?
I have used CUDA + KNN. Very powerful. You can also do feature selection with it. See https://www.researchgate.net/publication/262949628_Fast_Feature_Selection_in_a_GPU_Cluster_Using_the_Delta_Test?ev=prf_pub
Article Fast Feature Selection in a GPU Cluster Using the Delta Test
what is your data scale? my data is so huge, maybe cuad is not proper for it.. and could you give the source code link..
I just have published a paper that proposes a new distance metric which was invariant to the dimensionality of the feature vector, you may find the paper at the link:
https://www.researchgate.net/publication/264995324_Dimensionality_Invariant_Similarity_Measure?ev=prf_pub
Article Dimensionality Invariant Similarity Measure
Thank for all of you who reply my question. I have found code from http://users.eecs.northwestern.edu/~wkliao/Kmeans/
Parallel K-Means Data Clustering
it is working nice...
Hello Osman,
The number of samples for KNN classification is very dependent on the specific problem. For example, we use FLANN libraries in place recognition for matching scenes in different seasons of the year, and depending on the number of samples the performance and efficiency varies. I attach you our paper where we employ FLANN libraries.
Good luck with your research!
Conference Paper Towards Life-Long Visual Localization using an Efficient Mat...
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Algorithms