the structure/content of the python notebook are listed as follows:

part1:

• Features from pretrained models like VGG-16, VGG-19, ResNet-50, InceptionV3 and MobileNet

• Benchmark feature extraction speed with different model

• finetuned MobileNet with caltech101 labeled data

part2:

• Index feature list with Nearest Neighbor Brute Force Algorithm and visualize with PCA and t-SNE

part3:

• benchmark the algorithms based on the time it takes to index images and locate the most similar image based on its features using the Caltech-101 dataset.

TODO:

• Reduce feature dimension

• Query accuracy with Euclidean L2 vs cosine vs regular Euclidean distance

Some benchmarks on different algorithms to see relative speeds

These results lead to the benchmarking of time for indexing and searching on Caltech101. Repeating Level 2 on the Caltech256 features we can benchmark that as well.

Benchmarking the different models on Caltech101. (Rounded to nearest integer)

Algorithm	Number of features indexed	Time to search 1 image (ms)	Time to search 100 images (ms)	Time to search 1000 images (ms)	Time to index (ms)
Brute force	2048	14	38	240	22
k-d tree	2048	16	2270	24100	1020
Ball tree	2048	15	1690	17000	1090
PCA + brute force	100	1	13	135	0.334
PCA + k-d tree	100	1	77	801	20
PCA + ball tree	100	1	80	761	23
Annoy	2048	0.16	40	146	1420
PCA + Annoy	100	.008	2.3	20.3	109