Image-Sequential Binomial Self-Compensation (I-BSC)

• The paper introduces I-BSC, a method that applies binomial weighting to image sequences to suppress motion-induced ripple errors in dynamic PSP with exponential decay.
• It achieves computational efficiency by reducing phase-retrieval operations to a single arctangent evaluation per pixel, leading to significant speedup over phase-sequential approaches.
• Empirical results show error reductions from approximately 300 µm to 50 µm at practical compensation orders while maintaining full spatial resolution and camera-rate depth maps.

Searching arXiv for the specified BSC papers and related PSP motion-error literature. arxiv_search query: (Zhang et al., 2024) binomial self-compensation motion error dynamic 3d scanning Image-Sequential Binomial Self-Compensation (I-BSC) is a motion-error suppression method for dynamic phase-shifting profilometry (PSP) that applies binomial weighting directly to captured fringe images rather than to previously recovered phase maps. In the BSC line of work, the 2024 formulation established a pixel-wise and frame-wise loopable binomial self-compensation strategy for four-step PSP, and the 2025 formulation recharacterized that earlier approach as phase-sequential BSC (P-BSC) and introduced I-BSC as its image-sequential generalization (Zhang et al., 2024, Zhang et al., 14 Jul 2025). The defining claim of I-BSC is that successive motion-affected measurements can be combined with binomial coefficients so that the dominant ripple-shaped motion error decays exponentially with the compensation order $K$, while the wrapped phase is recovered with only a single arctangent evaluation per pixel (Zhang et al., 14 Jul 2025).

1. Genealogy within dynamic PSP

Dynamic PSP is ordinarily derived from an $N$-step phase-shifting sequence in which nominal phase shifts are $2\pi/N$ apart. In static measurement, PSP is favored for its high accuracy, robustness, and pixel-wise property, but dynamic measurement violates the assumption that the object remains static, leading to ripple-like errors in the reconstructed point clouds (Zhang et al., 2024). The 2024 binomial self-compensation paper addressed this by summing successive motion-affected phase frames with binomial coefficients in a pixel-wise and frame-wise loopable manner for the four-step case. The 2025 paper then distinguished that earlier formulation as P-BSC and proposed I-BSC, which transfers the binomial operation from phase sequences to image sequences and thereby computes the arctangent function only once (Zhang et al., 14 Jul 2025).

This shift from phase-sequential to image-sequential processing is the central conceptual move. P-BSC and I-BSC share the same compensatory principle—binomially weighted self-cancellation of motion-induced harmonic error—but they differ in where the weighting is applied. A plausible implication is that I-BSC should be understood less as a new error model than as a computational reformulation of the same suppression mechanism.

2. Motion-error model

In the dynamic PSP model summarized for I-BSC, the wrapped phase is estimated from a moving sequence by

$$\tilde{\phi}_i(x,y)=\arctan\!\left[\frac{\sum_{n=0}^{N-1} I_{i+n}(x,y)\sin(2\pi n/N)}{\sum_{n=0}^{N-1} I_{i+n}(x,y)\cos(2\pi n/N)}\right].$$

The captured intensity is written as

$$I_i(x,y)=A+B\cos\!\left[\phi_0(x,y)-\frac{2\pi i}{N}+x_i\right],$$

where $x_i$ is the unknown extra phase offset caused by motion. Linearization for small $x_i$ yields a dominant residual error

$$\epsilon(\tilde{\phi}_i)\simeq \frac{1}{N}\sum_{n=0}^{N-1} x_{i+n}\,[1-\cos(2\phi_{i+n})],$$

which decomposes into a DC term and a harmonic term at twice the fringe frequency. The practical 3D error is the high-frequency ripple, because the DC component simply shifts the whole cloud (Zhang et al., 14 Jul 2025).

The earlier four-step analysis makes the same structure explicit in a different notation. There, the raw motion-affected wrapped phase is expanded as

$\tilde{\phi}_i=\phi_i+\delta_i+H_i(\phi_i)+O(x^2),$

where $\delta_i$ is a DC lag term and

$N$0

is the ripple-shaped harmonic error (Zhang et al., 2024). Across both formulations, the error targeted by BSC is therefore not arbitrary dynamic distortion but specifically the motion-induced harmonic ripple that contaminates phase retrieval.

3. Image-sequential construction

The detailed I-BSC construction in the 2025 derivation assumes a cyclic $N$1 phase-shifting order. A total of $N$2 images is acquired:

$N$3

These are partitioned into four homogeneous-polarity streams,

$N$4

and each stream is binomially weighted:

$N$5

The wrapped phase at the central time index is then recovered by a single four-step call,

$N$6

The paper characterizes this as a 1-D convolution of each polarity stream with the $N$7th row of Pascal’s triangle (Zhang et al., 14 Jul 2025).

Operationally, the method proceeds by acquiring the cyclic $N$8-shifted sequence, accumulating the four compensated images, computing the single arctangent, and then sliding the window by one frame to produce a depth map for each new capture. This is the basis of the “frame-wise loopable” property. The 2024 predecessor achieved the same loopability through phase-map accumulation in a Yang Hui’s triangle pyramid; I-BSC preserves the loopable structure while moving the binomial accumulation earlier in the pipeline (Zhang et al., 2024).

4. Error-suppression mechanism

The mechanism of BSC is finite-difference cancellation. In the phase-sequential derivation, summing $N$9 successive motion-affected phase frames with binomial coefficients causes the harmonic terms to collapse into a $2\pi/N$0th finite difference of the motion offsets, multiplied by $2\pi/N$1. The resulting harmonic error amplitude therefore decays exponentially in $2\pi/N$2, and Fig. 2(b) in the 2024 paper confirms an approximately $2\pi/N$3 scaling (Zhang et al., 2024).

For I-BSC, the corresponding residual ripple is given as

$2\pi/N$4

where $2\pi/N$5 is the $2\pi/N$6th-order finite difference of the motion offsets. Because the binomial-weighted sum yields these high-order differences multiplied by $2\pi/N$7, the harmonic error decays at least as $2\pi/N$8. The paper further states that in practice one observes a slightly faster decay than pure $2\pi/N$9 because the higher-order differences themselves shrink for smooth motion (Zhang et al., 14 Jul 2025).

The significance of this result is structural rather than heuristic. I-BSC does not estimate motion explicitly, does not introduce an intermediate motion variable, and does not depend on an external motion model. Error compensation is produced by algebraic recombination of the motion-affected measurements themselves. This suggests that the method belongs to a class of self-canceling temporal filters whose efficacy depends on the smoothness of the motion-induced phase offsets.

5. Algorithmic and computational characteristics

The principal computational distinction between P-BSC and I-BSC is the number of phase-retrieval operations. P-BSC forms multiple phase maps and repeatedly combines them; I-BSC forms compensated images and performs one final phase retrieval. The 2025 paper summarizes the complexity contrast as follows (Zhang et al., 14 Jul 2025):

Method	Weighted object	Overall cost
P-BSC	Successive phase maps	$$\tilde{\phi}_i(x,y)=\arctan\!\left[\frac{\sum_{n=0}^{N-1} I_{i+n}(x,y)\sin(2\pi n/N)}{\sum_{n=0}^{N-1} I_{i+n}(x,y)\cos(2\pi n/N)}\right].$$0
I-BSC	Homogeneous fringe images	$$\tilde{\phi}_i(x,y)=\arctan\!\left[\frac{\sum_{n=0}^{N-1} I_{i+n}(x,y)\sin(2\pi n/N)}{\sum_{n=0}^{N-1} I_{i+n}(x,y)\cos(2\pi n/N)}\right].$$1

More specifically, P-BSC must form $$\tilde{\phi}_i(x,y)=\arctan\!\left[\frac{\sum_{n=0}^{N-1} I_{i+n}(x,y)\sin(2\pi n/N)}{\sum_{n=0}^{N-1} I_{i+n}(x,y)\cos(2\pi n/N)}\right].$$2 phase maps $$\tilde{\phi}_i(x,y)=\arctan\!\left[\frac{\sum_{n=0}^{N-1} I_{i+n}(x,y)\sin(2\pi n/N)}{\sum_{n=0}^{N-1} I_{i+n}(x,y)\cos(2\pi n/N)}\right].$$3, then perform $$\tilde{\phi}_i(x,y)=\arctan\!\left[\frac{\sum_{n=0}^{N-1} I_{i+n}(x,y)\sin(2\pi n/N)}{\sum_{n=0}^{N-1} I_{i+n}(x,y)\cos(2\pi n/N)}\right].$$4 additions/subtractions and $$\tilde{\phi}_i(x,y)=\arctan\!\left[\frac{\sum_{n=0}^{N-1} I_{i+n}(x,y)\sin(2\pi n/N)}{\sum_{n=0}^{N-1} I_{i+n}(x,y)\cos(2\pi n/N)}\right].$$5 arctan calls. I-BSC performs $$\tilde{\phi}_i(x,y)=\arctan\!\left[\frac{\sum_{n=0}^{N-1} I_{i+n}(x,y)\sin(2\pi n/N)}{\sum_{n=0}^{N-1} I_{i+n}(x,y)\cos(2\pi n/N)}\right].$$6 additions to build the compensated images, plus a single arctan per pixel, for overall cost $$\tilde{\phi}_i(x,y)=\arctan\!\left[\frac{\sum_{n=0}^{N-1} I_{i+n}(x,y)\sin(2\pi n/N)}{\sum_{n=0}^{N-1} I_{i+n}(x,y)\cos(2\pi n/N)}\right].$$7 (Zhang et al., 14 Jul 2025). The paper therefore describes I-BSC as reducing computational complexity by one polynomial order relative to P-BSC.

The concrete example given is a 640×480 setting with $$\tilde{\phi}_i(x,y)=\arctan\!\left[\frac{\sum_{n=0}^{N-1} I_{i+n}(x,y)\sin(2\pi n/N)}{\sum_{n=0}^{N-1} I_{i+n}(x,y)\cos(2\pi n/N)}\right].$$8, where the reduction translates into a 10×–20× speedup on a single CPU core. On a single-threaded 2.1 GHz i7, P-BSC ($$\tilde{\phi}_i(x,y)=\arctan\!\left[\frac{\sum_{n=0}^{N-1} I_{i+n}(x,y)\sin(2\pi n/N)}{\sum_{n=0}^{N-1} I_{i+n}(x,y)\cos(2\pi n/N)}\right].$$9) achieves approximately 5 fps, whereas I-BSC achieves approximately 50 fps, and the speedup grows roughly linearly with $$I_i(x,y)=A+B\cos\!\left[\phi_0(x,y)-\frac{2\pi i}{N}+x_i\right],$$0 (Zhang et al., 14 Jul 2025). This computational reformulation is one of the main reasons I-BSC is presented as a practical high-speed alternative to phase-sequential accumulation.

6. Empirical behavior, operating regime, and limitations

The reported empirical behavior is consistent with the finite-difference theory. In simulation with uniform acceleration, I-BSC’s RMS error falls by more than one half each time $$I_i(x,y)=A+B\cos\!\left[\phi_0(x,y)-\frac{2\pi i}{N}+x_i\right],$$1 increases by one. In real experiments on a moving plate, the motion ripple at $$I_i(x,y)=A+B\cos\!\left[\phi_0(x,y)-\frac{2\pi i}{N}+x_i\right],$$2 is reduced from $$I_i(x,y)=A+B\cos\!\left[\phi_0(x,y)-\frac{2\pi i}{N}+x_i\right],$$3 down to $$I_i(x,y)=A+B\cos\!\left[\phi_0(x,y)-\frac{2\pi i}{N}+x_i\right],$$4 (Zhang et al., 14 Jul 2025). The predecessor paper reported the same general trend for phase-sequential BSC: ripple errors of approximately $$I_i(x,y)=A+B\cos\!\left[\phi_0(x,y)-\frac{2\pi i}{N}+x_i\right],$$5 in raw four-step PSP, reduced to approximately $$I_i(x,y)=A+B\cos\!\left[\phi_0(x,y)-\frac{2\pi i}{N}+x_i\right],$$6 at $$I_i(x,y)=A+B\cos\!\left[\phi_0(x,y)-\frac{2\pi i}{N}+x_i\right],$$7 and approximately $$I_i(x,y)=A+B\cos\!\left[\phi_0(x,y)-\frac{2\pi i}{N}+x_i\right],$$8 at $$I_i(x,y)=A+B\cos\!\left[\phi_0(x,y)-\frac{2\pi i}{N}+x_i\right],$$9, with mean absolute error decaying roughly as $x_i$0 (Zhang et al., 2024).

The temporal claim attached to I-BSC is “quasi-single-shot” rather than literal single-shot. Because any new frame can be combined with the previous $x_i$1 to form a new depth map, the depth-map rate equals the camera rate, for example 90 fps, while spatial resolution remains full-pixel at 640×480 and temporal resolution is camera-limited (Zhang et al., 14 Jul 2025). This suggests that I-BSC should not be conflated with single-exposure profilometry: each estimate still depends on a sliding temporal window of $x_i$2 frames. The earlier BSC paper makes the latency explicit as approximately $x_i$3, with the example $x_i$4 for $x_i$5, $x_i$6, and it notes the need to buffer $x_i$7 frames and store $x_i$8 intermediate wrapped-phase layers in the phase-sequential implementation (Zhang et al., 2024).

The stated operating assumptions are equally important. I-BSC is derived under the small-motion approximation $x_i$9, $x_i$0 (Zhang et al., 14 Jul 2025). The earlier paper also assumes low-frequency object texture so that motion-induced phase offsets vary slowly in space; very high spatial texture may degrade the finite-difference smoothness (Zhang et al., 2024). Choice of $x_i$1 is a practical trade-off: $x_i$2–$x_i$3 is typically sufficient to suppress motion ripples to within sensor noise, whereas $x_i$4 gives diminishing returns because $x_i$5 is already small and image noise begins to dominate (Zhang et al., 14 Jul 2025). For noise, the effective variance empirically scales as approximately $x_i$6, whereas a straight $x_i$7-step PSP achieves approximately $x_i$8, so very large $x_i$9 is not economical for denoising (Zhang et al., 14 Jul 2025).

The main limitations are also explicit. Small projector-camera nonlinearity ($$\epsilon(\tilde{\phi}_i)\simeq \frac{1}{N}\sum_{n=0}^{N-1} x_{i+n}\,[1-\cos(2\phi_{i+n})],$$0) impairs harmonic cancellation; the recommended mitigation is full-chain nonlinearity rectification (OPWM + CNRC) first (Zhang et al., 14 Jul 2025). I-BSC corrects only the phase-shift error along the optical axis, and large off-axis motion still introduces artifacts; the 2025 paper identifies coupling with optical-flow tracking as future work (Zhang et al., 14 Jul 2025). The 2024 paper further notes that extension beyond $$\epsilon(\tilde{\phi}_i)\simeq \frac{1}{N}\sum_{n=0}^{N-1} x_{i+n}\,[1-\cos(2\phi_{i+n})],$$1 steps requires a divide-and-conquer phase-shift design, and it lists higher-order cascaded BSC, adaptive $$\epsilon(\tilde{\phi}_i)\simeq \frac{1}{N}\sum_{n=0}^{N-1} x_{i+n}\,[1-\cos(2\phi_{i+n})],$$2, a beam-splitter monocular setup, and combination with nonlinear error suppression such as binary defocus calibration as possible extensions (Zhang et al., 2024).

Taken together, these results place I-BSC as a specialized dynamic-PSP compensator with a sharply defined niche: it is pixel-wise, frame-wise loopable, algebraically self-compensating, and computationally lighter than P-BSC, but its guarantees rely on smooth motion offsets, calibrated nonlinearity, and the distinction between camera-rate output and multi-frame temporal support.