As [image-serch2], this time the train images's descriptors are read from the file,(Q: is it faster ?)
Yes, It is Faster, but My question is : the descriptors computing is not the bottleneck, the matching is. Why This can make it faster ???
SO, This time Use 1000 images as the dataset to see how query image size impact the retrieval score!
Scale Factor | Retrieval Score |
---|---|
1 | 3.156 |
0.95 | 3.152 |
0.9 | 3.152 |
0.85 | 3.124 |
0.8 | 3.056 |
0.75 | 3.04 |
0.7 | 3.04 |
0.65 | 2.944 |
0.6 | 2.884 |
0.55 | 2.812 |
0.5 | 2.728 |
0.45 | 2.6 |
0.4 | 2.452 |
0.35 | 2.252 |
0.3 | 2.024 |
0.25 | 1.612 |
0.2 | 0.2 |
<=0.1 | 0 |