In this question, the aim is to identify an "unknown" object depicted in a query image, by matching it to multiple reference images, and selecting the highest scoring

match. Since we only have one reference image per object, there is at most one correct answer. This is useful for example if you want to automatically identify a book

from a picture of its cover, or a painting or a geographic location from an unlabelled photograph of it$.$

The steps are as follows:

Select a set of reference images and their corresponding query images.

A$.$ Hint $1$: Start with the book covers, or just a subset of them.

B$.$ Hint $2$: This question can require a lot of computation to run from start to finish, so cache intermediate results $($e$.$g$.$ feature descriptors$)$ where you can.

Choose one query image corresponding to one of your reference images. Use RANSAC to match your query image to each reference image, and count the number

of inlier matches found in each case. This will be the matching score for that image.

Identify the query object. This is the identity of the reference image with the highest match score, or "not in dataset" if the maximum score is below a threshold.

Repeat steps $2-3$ for every query image and report the overall accuracy of your method $($that is$,$ the percentage of query images that were correctly matched in the

dataset$).$ Discussion of results should include both overall accuracy and individual failure cases.

A$.$ Hint $1$: In case of failure, what ranking did the actual match receive? If we used a "top$-$k$"$ accuracy measure, where a match is considered correct if it appears in

the top k match scores, would that change the result?

Your explanation of what you have done, and your results, here

Choose some extra query images of objects that do not occur in the reference dataset. Repeat step $4$ with these images added to your query set. Accuracy is now

measured by the percentage of query images correctly identified in the dataset, or correctly identified as not occurring in the dataset. Report how accuracy is

altered by including these queries, and any changes you have made to improve performance.

Your explanation of results and any changes made here

Repeat step $4$ and $5$ for at least one other set of reference images from museum$\_$paintings or landmarks, and compare the accuracy obtained. Analyse both your

overall result and individual image matches to diagnose where problems are occurring, and what you could do to improve performance. Test at least one of your

proposed improvements and report its effect on accuracy.