---
title: Berkeley-Adobe Perceptual Patch Similarity
url: https://www.emergentmind.com/topics/berkeley-adobe-perceptual-patch-similarity
type: topic
---

# Berkeley-Adobe Perceptual Patch Similarity

The Berkeley-Adobe Perceptual Patch Similarity (BAPPS) paradigm defines a rigorous protocol and dataset for benchmarking perceptual similarity metrics at the level of image patches, focusing on complex geometric and chromatic distortions that evade classic full-reference (FR) image quality assessment (IQA) approaches. By leveraging psychophysical ground truth—including both two-alternative forced choice (2AFC) and just-noticeable-difference (JND) methodologies—BAPPS enables the design and objective annotation of similarity metrics that explicitly target human vision performance. Subsequent developments such as the EDOKS metric have utilized BAPPS to demonstrate superiority over traditional and deep-learning-based FR metrics, particularly in scenarios characterized by local geometric perturbations and texture variation [2601.19680].

## 1. Structure and Construction of the BAPPS Dataset

BAPPS is a patch-centric dataset, derived from the RAISE1k collection of high-quality images. Its critical feature is the focus on localized (patch) distortions rather than holistic image manipulations, allowing granular assessment of perceptual similarity under conditions where structural and chromatic variations are decoupled from global context.

- **Patch Extraction**: Patches sampled at multiple resolutions from RAISE1k images.
- **Distortion Types**: Includes "traditional" distortions (Gaussian blur, additive noise, JPEG compression), deep-model artifacts (GAN perturbations, super-resolution), geometric warping, and advanced texture transformations.
- **Perceptual Study Design**:
    - *2AFC*: Triplet-based comparison involving a reference (x) and two distorted patches (x₁, x₂), judged by a five-person panel.
    - *JND*: Binary similarity/vs/difference assessment (1 s display each, 250 ms ISI) by three jurors; only unanimous responses define the "same" and "not same" reference pools.

These structural decisions yield a dataset particularly suited for empirical, statistically robust comparison of FR metrics under complex, real-world distortions.

## 2. Protocols for Human Perceptual Annotation

BAPPS employs strict psychophysical protocols to generate objective ground truth:

- **2AFC (Two-Alternative Forced Choice)**: Human judges are tasked with identifying which of two distorted patches is more similar to a reference. Agreement among a majority (out of five) defines ground truth, minimizing bias and variance.
- **JND (Just-Noticeable-Difference)**: Each of three judges views a reference and a single distorted patch and decides if they are the "same" or "not same." Only those pairs eliciting unanimous judgments are retained for the "same" and "not same" test subsets.

This dual-protocol annotation ensures that similarity metrics can be evaluated for both ordinal ranking (2AFC) and threshold-based detection (JND), anchoring quantitative metric validation in well-known human perception frameworks.

## 3. Methodological Innovations: The EDOKS Metric

The EDOKS metric combines two dissimilarity terms—texture and chromatic—into a transparent, white-box similarity index:

### Texture Dissimilarity via Earth Mover’s Distance (EMD)

- **Signature Extraction**:
    - Conversion of input RGB images X, Y to intensity, split into non-overlapping patches.
    - Application of a 24-filter Gabor bank (four scales × six orientations) to each patch.
    - Computation and normalization of Gabor energy matrices; clustering of these responses into k centroids (feature-space vectors).
    - Image signature is defined as S_X = (V_X, W_X), a mixture of k centroids and associated weights.

- **EMD Computation**:
    - Modeling S_X and S_Y as discrete weighted distributions in ℝ²⁴.
    - Ground (cost) distance is L₁; optimal transport (flow) f_{ij} yields the EMD:
      $$
      EMD(S_X,S_Y) = \frac{\sum_{i=1}^{n}\sum_{j=1}^{m} d(v_{x_i},v_{y_j})\,f_{ij}}{\sum_{i=1}^{n}\sum_{j=1}^{m} f_{ij}}
      $$
    - Returns zero for identically textured pairs, increasing with divergence.

### Chromatic Dissimilarity in Oklab Perceptual Color Space

- **Oklab Projection**:
    - Each RGB pixel is mapped into the Oklab (L,a,b) space using Ottosson’s perceptually uniform transformation.
    - Pairwise ΔE between two pixels: 
      $$
      \Delta E(p_1,p_2) = \sqrt{(L_1 - L_2)^2 + (a_1 - a_2)^2 + (b_1 - b_2)^2}
      $$
- **Chroma Dissimilarity Over Image**:
    - For images $\hat X, \hat Y$, the average per-pixel $\Delta E$:
      $$
      OK(\hat X,\hat Y) = \frac{1}{MN} \sum_{i=1}^M\sum_{j=1}^N \Delta E(\hat X_{i,j},\hat Y_{i,j})
      $$

### Combined EDOKS Index

- **Weighted Sum and Inverse Conversion**:
    $$
    EDOK(X,Y) = \alpha\,EMD(S_X,S_Y)\;+\;(1-\alpha)\,OK(\hat X,\hat Y),\quad \alpha\in[0,1]
    $$
    $$
    EDOKS(X,Y) = \frac1{EDOK(X,Y)\;+\;c}
    $$
    - Default: patch size $p=128$, $\alpha=0.5$, $c$ as machine epsilon.

An ablation study suggests optimal $\alpha\approx0.21$ for JND, highlighting the importance of chroma in certain perceptual discrimination tasks.

## 4. Quantitative Evaluation and Comparative Outcomes

Evaluation utilizes the BAPPS validation set (RAISE1k-derived patches) and considers the following:

- **2AFC Accuracy**: EDOKS ($\alpha$=0.5) achieves 72% alignment with human judges; texture-only (71%), chroma-only (69%).
- **JND Score Separation**: Mean EDOKS for "same"/"not same" = 12.57 / 17.22 (ratio 1.37), exceeding classical metrics.
- **Statistical Correlation**: Assessed on JND subsets using SROCC, KROCC, PLCC:

| Metric  | SROCC | KROCC | PLCC |
|---------|-------|-------|------|
| EDOKS   | 0.537 | 0.419 | 0.548 |
| PSNR    | 0.390 | 0.304 | 0.465 |
| SSIM    | 0.365 | 0.284 | 0.405 |
| DISTS   | 0.649 | 0.513 | 0.671 |
| LPIPS   | 0.633 | 0.499 | 0.640 |

EDOKS outperforms classical low-level FR metrics and approaches the correlation levels of recent deep-learning metrics, while providing interpretable attribution.

## 5. Visual Explanations and Interpretability

EDOKS uniquely offers pixel-wise (and patch-wise) heatmaps that decompose similarity attribution:

- **EMD Difference Maps**: Visualize mean Gabor response magnitude differences for geometric/texture shifts.
- **OK Heatmaps**: Per-pixel ΔE highlights subtle chromatic changes.
- **Combined Map**: Fuses both modalities for comprehensive visualization.

Example outputs include: 
- Focused detection of geometric changes in backgrounds, while ignoring undistorted objects (e.g., warped butterfly).
- Edge erosion highlighted by texture maps (EMD) and color inversion captured by OK in negated/blurred examples.
- Selective attribution to color regions in edge cases (road sign discrimination), outperforming luminance-based indices.

## 6. Implications and Directions for Future Research

BAPPS and EDOKS collectively demonstrate that robust, interpretable perceptual metrics can be constructed through explicit, dual-modality modeling of texture and color. The white-box approach contrasts with deep black-box methods by enabling the visualization of similarity judgments. 

Potential future directions include:
- Fine-tuning patch size ($p$) and filter/clustering schemes for improved granularity.
- Extension of the metric with explicit shape or shading descriptors.
- Multi-objective optimization of the combination parameter ($\alpha$) for domain-specific applications, possibly along Pareto fronts.

A plausible implication is that transparent, per-region visual explanations will become essential for the deployment of perceptual assessment algorithms in tasks requiring accountability and diagnostic precision.

**Reference**: All technical, experimental, and methodological content as well as quantitative performance results are from [2601.19680].

Source: https://www.emergentmind.com/topics/berkeley-adobe-perceptual-patch-similarity