Problems

1 / 1

Anchor Box Generation

Computer Vision

Medium

Object detectors like Faster R-CNN and SSD generate a dense set of predefined bounding boxes called anchors at every position on a feature grid. Each anchor serves as an initial guess that the network refines during training.

Given a square feature grid size, the original image size, a list of scales, and a list of aspect ratios, generate all anchor boxes in image coordinates.

Algorithm

Compute the stride (spacing between grid cells in image space):

stride = \frac{image\_size}{feature\_size}

For each grid cell (i, j), compute the center in image coordinates:

cx = (j + 0.5) \times stride \qquad cy = (i + 0.5) \times stride

For each combination of scale s and aspect ratio r, compute the box width and height:

w = s \cdot \sqrt{r} \qquad h = \frac{s}{\sqrt{r}}

The anchor box is [cx - w/2, cy - h/2, cx + w/2, cy + h/2].

Iterate over grid cells in row-major order (i then j), and for each cell iterate over scales then aspect ratios.

Loading visualization...

Examples

Input: feature_size = 1, image_size = 8, scales = [4], aspect_ratios = [1.0]

Output: [[2.0, 2.0, 6.0, 6.0]]

Explanation: The only grid cell is centered at (4, 4), and its anchor has width and height 4.

Input: feature_size = 2, image_size = 8, scales = [2], aspect_ratios = [1.0]

Output: [[1.0, 1.0, 3.0, 3.0], [5.0, 1.0, 7.0, 3.0], [1.0, 5.0, 3.0, 7.0], [5.0, 5.0, 7.0, 7.0]]

Hint 1

Compute stride as image_size / feature_size and offset each grid coordinate by 0.5 before scaling.

Hint 2

For scale s and ratio r, use width = s * sqrt(r) and height = s / sqrt(r).

Requirements

Map each grid cell center to image coordinates using the 0.5 offset
Generate one anchor per (scale, aspect_ratio) pair at each grid position
Return anchors in row-major grid order, then by scale, then by ratio

Constraints

feature_size >= 1, image_size >= 1
scales and aspect_ratios are non-empty lists of positive floats
Return a list of [x1, y1, x2, y2] boxes as floats
Time limit: 300 ms

Try Similar Problems

Iou Bounding Box Non Maximum Suppression Roi Pooling Bilinear Interpolation Max Pooling 2d

Prev Next

Anchor Box Generation

Computer Vision

Medium

Given a square feature grid size, the original image size, a list of scales, and a list of aspect ratios, generate all anchor boxes in image coordinates.

Algorithm

Compute the stride (spacing between grid cells in image space):

stride = \frac{image\_size}{feature\_size}

For each grid cell (i, j), compute the center in image coordinates:

cx = (j + 0.5) \times stride \qquad cy = (i + 0.5) \times stride

For each combination of scale s and aspect ratio r, compute the box width and height:

w = s \cdot \sqrt{r} \qquad h = \frac{s}{\sqrt{r}}

The anchor box is [cx - w/2, cy - h/2, cx + w/2, cy + h/2].

Iterate over grid cells in row-major order (i then j), and for each cell iterate over scales then aspect ratios.

Loading visualization...

Examples

Input: feature_size = 1, image_size = 8, scales = [4], aspect_ratios = [1.0]

Output: [[2.0, 2.0, 6.0, 6.0]]

Explanation: The only grid cell is centered at (4, 4), and its anchor has width and height 4.

Input: feature_size = 2, image_size = 8, scales = [2], aspect_ratios = [1.0]

Output: [[1.0, 1.0, 3.0, 3.0], [5.0, 1.0, 7.0, 3.0], [1.0, 5.0, 3.0, 7.0], [5.0, 5.0, 7.0, 7.0]]

Hint 1

Compute stride as image_size / feature_size and offset each grid coordinate by 0.5 before scaling.

Hint 2

For scale s and ratio r, use width = s * sqrt(r) and height = s / sqrt(r).

Requirements

Map each grid cell center to image coordinates using the 0.5 offset
Generate one anchor per (scale, aspect_ratio) pair at each grid position
Return anchors in row-major grid order, then by scale, then by ratio

Constraints

feature_size >= 1, image_size >= 1
scales and aspect_ratios are non-empty lists of positive floats
Return a list of [x1, y1, x2, y2] boxes as floats
Time limit: 300 ms

Try Similar Problems

Iou Bounding Box Non Maximum Suppression Roi Pooling Bilinear Interpolation Max Pooling 2d

Case 1

Case 2

feature_size =

Accepts: number

image_size =

Accepts: number

scales =

Accepts: array

aspect_ratios =

Accepts: array

Given a square feature grid size, the original image size, a list of scales, and a list of aspect ratios, generate all anchor boxes in image coordinates.

Algorithm

Compute the stride (spacing between grid cells in image space):

stride = \frac{image\_size}{feature\_size}

For each grid cell (i, j), compute the center in image coordinates:

cx = (j + 0.5) \times stride \qquad cy = (i + 0.5) \times stride

For each combination of scale s and aspect ratio r, compute the box width and height:

w = s \cdot \sqrt{r} \qquad h = \frac{s}{\sqrt{r}}

The anchor box is [cx - w/2, cy - h/2, cx + w/2, cy + h/2].

Iterate over grid cells in row-major order (i then j), and for each cell iterate over scales then aspect ratios.