--- title: Continuous Cost Aggregation (CCA) url: https://www.emergentmind.com/topics/continuous-cost-aggregation-cca type: topic --- # Continuous Cost Aggregation (CCA) Continuous Cost Aggregation (CCA) is an algorithmic framework for extracting continuous, subpixel disparities from Dual-Pixel (DP) sensor data, leveraging local parabolic modeling, path-wise quadratic aggregation, and multi-scale coefficient fusion. CCA was introduced to address the challenge posed by DP images’ tiny baseline and non-uniform point spread function (PSF), which preclude conventional stereo matching algorithms from yielding accurate depth information. CCA combines closed-form subpixel disparity estimation within a semi-global matching (SGM) paradigm with efficient propagation of quadratic cost coefficients, enabling pixel-wise minimization without discrete winner-take-all steps and substantially reduced memory requirements relative to classical cost-volume approaches [2306.07921]. ## 1. Local Parabolic Modeling of Pixelwise Matching Cost CCA begins with rectified dual-pixel images $I_L,I_R$ and computes a discrete per-pixel cost volume $C_{\text{int}}(p,d)$ for integer disparities $d\in[d_{\min},\dots,d_{\max}]$, typically via metrics such as Sum of Absolute Differences (SAD) or Normalized Cross-Correlation (NCC). Due to the minute DP baseline, subpixel accuracy is essential. Thus, for each pixel $p$, a parabola is fit locally around the integer cost minimum: - Identify $d^0_p = \arg\min_n C_{\text{int}}(p,n)$ - Fit $C_{p,d^0_p}(\Delta d) = a_p \Delta d^2 + b_p \Delta d + c_p$ using costs at $d^0_p-1$, $d^0_p$, $d^0_p+1$ - Coefficient calculation: - $a_p = [C_{\text{int}}(p,d^0_p+1) + C_{\text{int}}(p,d^0_p-1) - 2C_{\text{int}}(p,d^0_p)] / 2$ - $b_p = [C_{\text{int}}(p,d^0_p+1) - C_{\text{int}}(p,d^0_p-1)] / 2$ - $c_p = C_{\text{int}}(p,d^0_p)$ - Transform back to global disparity $d=d^0_p+\Delta d$ to produce $C_p(d) = \alpha_p d^2 + \beta_p d + \gamma_p$ where - $\alpha_p = a_p$ - $\beta_p = b_p - 2a_p d^0_p$ - $\gamma_p = c_p + a_p (d^0_p)^2 - b_p d^0_p$ These parabolic models admit unique minima $d_p^{\text{loc}} = -\beta_p/(2\alpha_p)$ and allow the curvature $\alpha_p$ to serve as a confidence measure, with flatter parabolas used in ambiguous regions by scaling $\alpha_p$. ## 2. Path-wise Quadratic Aggregation and Smoothness Constraint CCA applies semi-global matching by aggregating quadratic cost functions along $R$ scanline directions. For each path $r$, the propagated cost is expressed as a quadratic $L_p^r(d)$, combining the local parabola and a quadratic smoothness penalty. At pixel $p$: - Let the predecessor’s optimum $m_{p-1} = \arg\min_d L_{p-1}^r(d)$ - Aggregate: $L_p^r(d) = C_p(d) + P_{\text{adapt}}(d-m_{p-1})^2$ - The coefficients update as: - $A_p^r = \alpha_p + P_{\text{adapt}}$ - $B_p^r = \beta_p + P_{\text{adapt}} \cdot (B_{p-1}^r/A_{p-1}^r)$ - $\Gamma_p^r = \gamma_p + P_{\text{adapt}} (B_{p-1}^r/(2A_{p-1}^r))^2$ - $m_{p-1} = -B_{p-1}^r / (2A_{p-1}^r)$ The adaptive smoothness weight $P_{\text{adapt}} = P A_{p-1}^r \exp[-(I_p - I_{p-1})^2 / \sigma^2]$ adjusts smoothing strength according to local image gradients, preserving discontinuities at strong edges. ## 3. Closed-form Subpixel Disparity Extraction After aggregation over all $R$ directions, each pixel $p$ has $R$ quadratic costs $\{A_p^r, B_p^r, \Gamma_p^r\}$. Summation yields a total quadratic cost: $$ S_p(d) = \sum_{r=1}^R L_p^r(d) = \left(\sum A_p^r \right) d^2 + \left(\sum B_p^r \right) d + \left(\sum \Gamma_p^r \right) $$ The subpixel disparity is extracted in closed form by solving for the minimum: $$ d_p^* = -\left(\sum_r B_p^r\right) / \left(2 \sum_r A_p^r\right) $$ This direct coefficient inversion eliminates the need for discrete label selection and enables efficient pixel-wise minimization. ## 4. Multi-scale Aggregation and Pyramid Fusion CCA implements a multi-scale strategy to enhance robustness against depth-dependent defocus blur. Using an image pyramid with scales $s = S,\dots,1$: 1. Run CCA at coarsest scale $S$ to obtain coefficient maps $A_{p,S}$, $B_{p,S}$. 2. Upsample priors for scale $S-1$: - $B_{p,S-1}^{\text{prior}} = \text{Upsample}(B_{p,S} \cdot F)$ - $A_{p,S-1}^{\text{prior}} = \text{Upsample}(A_{p,S} / F^2)$ 3. Inject priors with weight $w$ before fine-scale CCA: - $\alpha_{p,S-1} \leftarrow \alpha_{p,S-1} + w \cdot A_{p,S-1}^{\text{prior}}$ - $\beta_{p,S-1} \leftarrow \beta_{p,S-1} + w \cdot B_{p,S-1}^{\text{prior}}$ 4. Re-run CCA down to full resolution. In areas with weak texture or blur, the coarse-scale belief steers aggregation; in regions with high local confidence, fine-scale data predominate. ## 5. Algorithmic Structure CCA proceeds as follows: ```python # Inputs: I_L, I_R, D = {d_min...d_max}, R directions, S scales, T_s iterations, P, σ, w # 1. build image pyramid I^{(s)}, s=1...S for s = S down to 1: # 2. compute sparse cost-volume -> local parabolas for each pixel p: compute C_int(p,d) for d in D find d⁰_p = argmin_d C_int(p,d) fit (a_p,b_p,c_p) via 3-point parabola form (α_p,β_p,γ_p) if s1: A_prior = Upsample(A_p)/F^2 B_prior = Upsample(B_p)*F # 6. optional post-filter (left-right consistency, speckle removal) ``` ## 6. Experimental Protocols and Quantitative Results CCA has been quantitatively evaluated on several datasets: - **DSLR:** Canon DP dataset (Punnappurath et al., ICCP 2020) - **Phone:** Google Pixel 2/3 (Garg et al., ICCV 2019) - **Standard Stereo:** Middlebury 2014 (quarter-resolution) Metrics include affine-invariant errors (AI(1), AI(2)), $1-|ρ_s|$, bad pixel rates at thresholds 0.5, 1, 2 pixels, and RMSE. Key results below: **Table 1. DSLR Results (geometric mean of [AI(1),AI(2),1−|ρ_s|], lower is better)** | Method | AI(1) | AI(2) | 1−|ρ_s| | Geo. Mean | |------------------------|-------|-------|-------|-----------| | SDoF | 0.087 | 0.129 | 0.291 | 0.144 | | DPdisp | 0.047 | 0.074 | 0.082 | 0.065 | | DPE | 0.061 | 0.098 | 0.103 | 0.110 | | CCA | 0.041 | 0.068 | 0.061 | 0.053 | | CCA + filter | 0.036 | 0.061 | 0.049 | 0.048 | **Table 2. Phone (Pixel) Results (same metrics, lower is better)** | Method | AI(1) | AI(2) | 1−|ρ_s| | Geo. Mean | |---------------|-------|-------|-------|-----------| | SDoF | 0.027 | 0.037 | 0.236 | 0.063 | | CCA | 0.026 | 0.036 | 0.225 | 0.059 | | CCA + filter | 0.025 | 0.035 | 0.217 | 0.057 | **Table 3. Middlebury ¼-res Results (non-occluded, lower is better)** | Method | bad<0.5 px | bad<1 px | bad<2 px | RMSE | |----------------|------------|----------|----------|-------| | SGM | 26.1 % | 17.2 % | 12.2 % | 9.90 | | CCA | 26.2 % | 18.3 % | 13.2 % | 5.20 | | SGM + filter | 23.5 % | 15.2 % | 10.5 % | 4.04 | | CCA + filter | 24.6 % | 16.7 % | 11.6 % | 4.04 | CCA delivers continuous, subpixel disparities in closed form and operates with $\mathcal{O}(WHR + WHD)$ time and $\mathcal{O}(WH)$ memory, circumventing the need for full cost-volume storage. On DP images, the method surpasses prior learning and non-learning baselines; on standard stereo, it is comparable to Semi-Global Matching (SGM), with quantitative advantages in memory efficiency. ## 7. Context and Implications CCA addresses specific limitations of DP disparity extraction, namely sensitivity to PSF variation and the infeasibility of conventional stereo matching due to small disparities. The quadratic form persists throughout coefficient propagation, enabling closed-form minimization and robust multi-scale fusion. A plausible implication is that CCA’s framework could generalize to other continuous-label matching tasks where cost functions are locally convex and aggregatable under quadratic constraints. Its memory and time efficiency offer practical advantages for embedded or real-time applications. CCA’s competitive performance on standard stereo benchmarks (Middlebury) with a reduced computational footprint demonstrates its potential for broader adoption across passive depth sensing modalities [2306.07921]. Source: https://www.emergentmind.com/topics/continuous-cost-aggregation-cca