What is the highest possible value the mean Average Precision (mAP) can take?

Suppose:

1. we are computing the Bag of Visual Words (BoVW) vector from an image;
2. we computed 1517 descriptors (represented each by 128 integers) for this image;
3. our codebook contains 256 elements;
4. each element in our BoVW vector is represented by a float32 (4 bytes).

What is the size in bytes of the BoVW vector for this image (right after the encoding step, with the simplest possible pooling)?

This question takes several minutes. It requires you to plot an AP curve manually.

What would be the Average Precision (AP) @ 10 of the following query results, assuming that all the expected elements were retrieved:

1. relevant
2. relevant
3. NOT relevant
4. NOT relevant

5. NOT relevant
6. relevant
7. NOT relevant
8. NOT relevant
9. NOT relevant
10. NOT relevant
11. NOT relevant
12. NOT relevant

You can compute an approximation of the AP using the upper envelope of the Precision Recall curve.

Please provide at least 2 significant digits ("0.XY" format).

What is the lowest possible value the mean Average Precision (mAP) can take?

When plotting the Precision/Recall curve @ k for some query , what are the possible values for the y axis (aka vertical axis) under the following constraints?

Constraints:

• Our dataset contains 5000 elements.
• We consider only the first 10 elements for each query ().
• For this particular query , there are 5 equally-relevant items in our dataset.

When plotting the Precision/Recall curve @ k for some query , what are the possible values for the x axis (aka horizontal axis) under the following constraints?

Constraints:

• Our dataset contains 5000 elements.
• We consider only the first 10 elements for each query ().
• For this particular query , there are 5 equally-relevant items in our dataset.

Suppose we have a prototype CBIR pipeline which allows us to index global image descriptors using a linear indexing.

We measured a mean average precision (mAP) of for our system, using a test set T1.

We now plan to use another indexing technique suitable for a larger scale: we will use some approximate nearest neighbor technique (like hierarchical k-means or LSH).

What should be the expected, likely values for the mAP of our system with the new indexing approach, when using the same test set T1?

During the practice session 4, it is possible that some student get different quality (mAP) values while having the exact same code.

What can be the reason(s)?

Among the following steps, which ones are specific the content-based image retrieval (CBIR) pipeline, and do not belong to the image matching pipeline based on local features?

Both pipelines share some common step, but which ones belong exclusively to the CBIR pipeline?