Computer Vision Foundations · free preview

Why Multiscale

Perspective makes objects appear at many sizes, and multiscale analysis handles all of them with one small template.

1. One scene, many sizes

One scene, many sizes

2. Resizing templates is expensive

Resizing templates is expensive

3. Flip the resizing onto the image

Flip the resizing onto the image

4. Scale as a coordinate

Scale as a coordinate

5. Shrinking instead of stretching

A detector uses a 20×2020 \times 20 pixel bird template and wants to find birds four times larger in size. Why is downsampling the image by a factor of 4 and reusing the template cheaper than stretching the template to 80×8080 \times 80?

6. Reading a detection's scale

A pyramid shrinks linear size by a factor of 3/43/4 per level. A bird template matches at level 0, level 1 and level 3 of the same tree photo. Which birds are the largest in the original image?

7. Four levels of shrinking

A pyramid shrinks linear image size by a factor of 3/43/4 at each level. Starting from level 0 at full size, compute the linear scale of level 4 relative to the original, that is (3/4)4(3/4)^4, and write it as a two-decimal number.

Practice this interactively

Free account · instant grading · spaced review that schedules itself.

Start this course for free