README.md

metadata

title: Layout Occlusion
emoji: 🌍
colorFrom: blue
colorTo: gray
sdk: gradio
sdk_version: 4.36.1
app_file: app.py
pinned: false

Layout Occlusion

Description

The Layout Occlusion metric evaluates how much layout elements occlude or cover important visual regions in the background canvas. This metric is particularly important for content-aware layout generation where background imagery should remain visible and not be blocked by poorly placed elements.

What It Measures

This metric computes the average saliency (visual importance) of canvas regions covered by layout elements:

• Visual importance coverage: How much salient (visually important) content is blocked by elements
• Element placement quality: Whether elements are placed on less important background regions
• Content-awareness: How well the layout respects the underlying visual content

Lower occlusion scores indicate better placement where elements avoid covering important background content.

Metric Details

• Uses saliency maps to identify visually important regions in the canvas
• Computes average saliency values in areas covered by elements
• Combines two saliency maps for robust evaluation
• From PosterLayout (Hsu et al., CVPR 2023) for content-aware poster design
• Lower scores mean elements are placed on less salient (less important) regions

Usage

Installation

pip install evaluate

Basic Example

import evaluate
import numpy as np

# Load the metric with canvas dimensions
metric = evaluate.load(
    "creative-graphic-design/layout-occlusion",
    canvas_width=360,
    canvas_height=504
)

# Prepare data
predictions = np.random.rand(1, 25, 4)
gold_labels = np.random.randint(0, 4, size=(1, 25))
# Paths to saliency map images (grayscale, 0-255)
saliency_maps_1 = ["path/to/saliency_map_1.png"]
saliency_maps_2 = ["path/to/saliency_map_2.png"]

score = metric.compute(
    predictions=predictions,
    gold_labels=gold_labels,
    saliency_maps_1=saliency_maps_1,
    saliency_maps_2=saliency_maps_2
)
print(score)

Batch Processing Example

import evaluate

# Load the metric
metric = evaluate.load(
    "creative-graphic-design/layout-occlusion",
    canvas_width=360,
    canvas_height=504
)

# Batch processing
batch_size = 128
predictions = np.random.rand(batch_size, 25, 4)
gold_labels = np.random.randint(0, 4, size=(batch_size, 25))
saliency_maps_1 = [f"path/to/saliency_{i}_1.png" for i in range(batch_size)]
saliency_maps_2 = [f"path/to/saliency_{i}_2.png" for i in range(batch_size)]

score = metric.compute(
    predictions=predictions,
    gold_labels=gold_labels,
    saliency_maps_1=saliency_maps_1,
    saliency_maps_2=saliency_maps_2
)
print(score)

Parameters

Initialization Parameters

• canvas_width (int, required): Width of the canvas in pixels
• canvas_height (int, required): Height of the canvas in pixels

Computation Parameters

• predictions (list of lists of float): Normalized bounding boxes in ltrb format
• gold_labels (list of lists of int): Class labels for each element (0 = padding)
• saliency_maps_1 (list of str): File paths to first set of saliency map images
• saliency_maps_2 (list of str): File paths to second set of saliency map images

Note: Saliency maps should be grayscale images (0-255) where brighter regions indicate more visually important areas. They will be automatically resized to match canvas dimensions if needed.

Returns

Returns a float value representing the average saliency of occluded regions (range: 0.0 to 1.0).

Interpretation

• Lower is better (range: 0.0 to 1.0)
• Value ~0.0: Elements placed on unimportant background regions (ideal)
• Value ~0.5: Elements partially cover moderately important regions
• Value ~1.0: Elements heavily occlude highly salient background content (problematic)

Use Cases

• Content-aware layout generation: Evaluate if generated layouts respect background imagery
• Poster/flyer design: Ensure text and graphics don't block important visual elements
• Advertisement layout: Place call-to-action elements without covering key visuals
• Magazine/presentation layouts: Balance element placement with background content

Key Insights

• Good layouts minimize occlusion of salient background regions
• Background-aware models should achieve lower occlusion scores
• Trade-off: Sometimes covering salient regions is necessary for design needs
• Use with other metrics: Combine with validity and alignment for comprehensive evaluation

Citations

@inproceedings{hsu2023posterlayout,
  title={Posterlayout: A new benchmark and approach for content-aware visual-textual presentation layout},
  author={Hsu, Hsiao Yuan and He, Xiangteng and Peng, Yuxin and Kong, Hao and Zhang, Qing},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={6018--6026},
  year={2023}
}

References

• Paper: PosterLayout (Hsu et al., CVPR 2023)
• Reference Implementation: PosterLayout eval.py

Related Metrics

• Layout Utility: Measures how well suitable space is utilized
• Layout Unreadability: Evaluates text placement on non-flat regions
• Layout Validity: Checks basic validity constraints