Hampi Temples are spread over a wide region unlike the
Golkonda Fort. We collected images from dierent temples
and built a dataset of 5,718 images. This dataset is largely
comprised of images covering temple overviews and structures
and does not encompass much of the internal architecture,
which would be a much denser dataset. Similar to experiments
at Golkonda, we choose 10 important monuments
at dierent temples of Hampi to evaluate the Annotation
Accuracy on phone-captured query images. We achieve an
overall 93% mean Annotation Accuracy.
In another experiment at the Hazara Rama Temple, Hampi,
we collected an image dataset covering the internal architectural
details. This includes images of numerous sculpted
friezes depicting the ancient Indian epic, Ramayana. These
relief structures, stone-carved on the walls of the temple particularly
around the main shrine present the epic story in the
form of distinctive scenes(See Figure 9(a) ). We nd that the
previous vocabulary built on the 5.7K dataset is unsuitable
for these images, as there are fewer keypoints owing to the
distinct texture of the stone carvings. We compute a new
vocabulary from these images and build a BoW retrieval
system. Our application tries to identify and annotate each
of these scenes, when captured using a mobile phone. In
this way, a tourist can guide himself through the interesting
mythological story of Ramayana at this 15th century shrine
(See Figure 9(b) ).