Object detection¶

from annotools import (
    BBoxObject,
    GridOptions,
    draw_bboxes,
    draw_grid,
    encode,
    load_image,
    normalize_coordinates,
    preview,
)

view = preview(load_image("photo.jpg"), max_width=768, max_height=768)
view.image = draw_grid(view.image, GridOptions(columns=10, rows=10)).image
gridded = encode(view.image, "jpeg")  # ask the model for boxes on this view

width, height = view.metadata["output_size"]
answer = [[144, 72, 432, 360]]  # the model's pixels of the view it saw
boxes = normalize_coordinates(answer, width, height, crop=view.metadata["crop"])

check = draw_bboxes(view, [BBoxObject(bbox=boxes[0], label="cat")])

boxes is [[0.1875, 0.125, 0.5625, 0.625]] — normalized to the uncropped source, which is what the store keeps. draw_grid returns a new result, but only its image is taken: view keeps the crop and scale that both normalize_coordinates and draw_bboxes map through, so the boxes land on the same gridded view the model answered on.

preview_image_grid, then normalize_coordinates on the answer, then preview_image_bboxes to show the model its own boxes:

{ "source": "photo.jpg", "columns": 10, "rows": 10, "max_width": 768, "max_height": 768 }

{
  "coordinates": [[144, 72, 432, 360]],
  "base_width": 768,
  "base_height": 576,
  "crop": [0, 0, 1, 1]
}

{
  "source": "photo.jpg",
  "objects": [{ "bbox": [0.1875, 0.125, 0.5625, 0.625], "label": "cat" }],
  "grid": { "columns": 10, "rows": 10 },
  "max_width": 768,
  "max_height": 768
}

base_width / base_height and crop are the grid call's own output_width, output_height and applied crop — read them off its metadata rather than assuming, since the fit depends on the source aspect ratio (768x576 for a 4:3 photo) and crop is only the full frame while the model is looking at the whole image.

The grid call adds grid (columns, rows, step_x / step_y, cell_width / cell_height) to the usual metadata; normalize_coordinates answers {"coordinates": [[...]]}; the box call adds objects, the number drawn. Repeat propose → verify → correct by index for a bounded number of rounds, then commit each accepted box to the store.