CPICD

CPICD provides recalibrated ground-truth crop boxes for aesthetic image cropping on GAIC, FLMS, and FCDB. These annotations correspond to the CPICD evaluation in Discrete Annotation, Continuous Preference: Rethinking Supervision for Accurate and Generalizable Aesthetic Image Cropping.

Project: https://github.com/zzqingz/CPIC

Files

File	Images	Crop boxes
GAIC.json	200	469
FLMS.json	500	1,873
FCDB.json	345	345
Total	1,045	2,687

This package contains annotations only. Obtain the corresponding images from the original GAIC, FLMS, and FCDB datasets and match them using image_id. Each file contains the evaluation images used for the corresponding CPICD subset; it does not include a training split.

CPICD300.json is a 300-image subset (100 randomly sampled from each dataset, with a dataset field identifying the source) provided for convenient evaluation.

Original images

Download the original images from the following sources and match them to the annotations using image_id (and dataset for CPICD300).

Dataset	Image download
GAIC	Official repository — Dataset link
FLMS	Original image archive (HTTP)
FCDB	Author-provided image archive · Official repository

Use the CPICD JSON files supplied here as the evaluation annotations.

Annotation format

Each per-dataset JSON file is a list of image records with exactly four fields:

Field	Description
image_id	Original image filename, including its extension. Unique within each JSON file.
width	Image width in pixels.
height	Image height in pixels.
boxes	A list of valid crop boxes, each stored as [x1, y1, x2, y2].

Coordinates are floating-point pixel coordinates in the original image: the origin is at the top-left, x increases to the right, and y increases downward. Every box satisfies 0 <= x1 < x2 <= width and 0 <= y1 < y2 <= height. Preserve the supplied precision during evaluation.

All three files use the same xyxy convention, including GAIC and FCDB. The boxes are not row/column coordinates, xywh coordinates, or normalized coordinates.

Each image has one to four valid crops. Boxes are evaluated as a set; their array positions do not represent quality ranks. The annotations contain no scores or weights.

Example from GAIC.json:

{
  "image_id": "210333.jpg",
  "width": 1024,
  "height": 678,
  "boxes": [
    [
      70.5,
      0.0,
      1024.0,
      557.3
    ],
    [
      143.4,
      0.0,
      912.0,
      589.1
    ]
  ]
}

Use the pair (dataset, image_id) when combining the three files.