---
title: 'CAMERA-P: Multi-Modal Imaging and Compression'
url: https://www.emergentmind.com/topics/camera-p
type: topic
---

# CAMERA-P: Multi-Modal Imaging and Compression

Searching arXiv for recent and exact matches on “CAMERA-P” and the provided IDs to ground the article.
arxiv_search(query="CAMERA-P OR pCam6060 OR \"Disturbance-free single-pixel imaging camera\" OR \"Triple Coding Empowered FDMA-CDMA Mode High Security CAOS Camera\" OR \"CAMERA: Multi-Matrix Joint Compression\" OR \"Cherenkov Camera for the PBR mission\"", max_results=10, sort_by="submittedDate") > /tmp/arxiv_results.json && cat /tmp/arxiv_results.json

I found relevant arXiv records matching the provided usages of “CAMERA-P,” including the single-pixel imaging work, CAOS FDMA-CDMA mode, the pCam6060 photodetecting system, the MoE compression framework, and PBR Cherenkov-camera papers.
CAMERA-P is a context-dependent designation used in several technically unrelated research literatures. In the available record, it can denote the structured pruning component of the CAMERA framework for mixture-of-experts compression, the pCam6060 photodetecting system optimized for astronomical photometry, a P-channel FDMA-CDMA operating mode of the CAOS camera, a disturbance-suppressed single-pixel imaging configuration associated with \(P=0.5\) Bernoulli modulation, and, in some query conventions, fast SiPM-based Cherenkov camera packages for balloon-borne or ground-based high-energy astrophysics [2508.02322], [2412.10109], [2107.06175], [2204.07711], [2511.15636], [2406.08274]. This suggests that CAMERA-P is best understood as a disambiguation term whose precise referent depends on subfield and paper-specific nomenclature.

## 1. Scope and disambiguation

The term does not identify a single canonical instrument or algorithm. One paper explicitly defines **CAMERA-P** as a structured pruning method inside a broader MoE-compression framework. Other papers use closely related but not identical labels: one describes CAMERA-P as the **pCam6060** astronomical photodetecting system, another explains that CAMERA-P is effectively the **CAOS camera operated in a “P-channel” FDMA-CDMA configuration**, and a PBR proceedings contribution notes that the query term appears to refer to the mission’s **Cherenkov Camera (CC)** rather than to an independent formal acronym. In the supplied single-pixel-imaging mapping, CAMERA-P is associated with the **CSPI** camera based on complementary detection and optimized Bernoulli modulation [2508.02322], [2412.10109], [2107.06175], [2204.07711], [2511.15636], [2406.08274].

| Context | Referent | Core technical idea |
|---|---|---|
| MoE compression | CAMERA-P | Structured micro-expert pruning |
| Coded optical sensing | P-channel CAOS camera | FDMA-CDMA triple coding with \(P\) frequency channels |
| Single-pixel imaging | CSPI / CAMERA-P mapping | Complementary detection with \(P=0.5\) Bernoulli modulation |
| Astronomical photometry | pCam6060 / CAMERA-P | Large-format BSI CMOS HDR photodetecting system |
| High-energy astrophysics | PBR CC or modular SiPM camera | Fast Cherenkov detection with SiPM focal planes |

This distribution matters because the meaning of the suffix **“-P”** changes across papers. In the MoE work it names a specific pruning operator; in the CAOS work it denotes the number of simultaneous FDMA channels; in pCam6060 it is part of the camera name; and in mission instrumentation papers it is an external query label mapped onto camera subsystems rather than the instrument’s own primary acronym.

## 2. CAMERA-P as structured micro-expert pruning in MoE models

The most explicit formal definition appears in **“CAMERA: Multi-Matrix Joint Compression for MoE Models via Micro-Expert Redundancy Analysis”**, where CAMERA-P is the structured pruning instantiation of the CAMERA framework [2508.02322]. The paper’s central move is to replace the usual compression unit—expert, matrix, or channel—with the **micro-expert**, defined as the coupled computation associated with one hidden dimension across the expert FFN’s up, gate, and down projections. In the paper’s notation,
$$
e_i(\mathbf{x}) = \mathbf{w}_i^{\mathrm{down}}\!\left[\sigma\!\left(\mathbf{w}_i^{\mathrm{gate}}\mathbf{x}\right)\cdot \mathbf{w}_i^{\mathrm{up}}\mathbf{x}\right].
$$
This induces a decomposition of the MoE layer as a mixture of micro-experts,
$$
\mathbf{y} = \sum_{i=1}^{N_e} \phi_i \mathbf{w}_i^{\mathrm{down}}, \quad
\phi_i = A_i(\mathbf{x})\cdot \sigma(\mathbf{w}_i^{\mathrm{gate}}\mathbf{x})\cdot \mathbf{w}_i^{\mathrm{up}}\mathbf{x}.
$$

The redundancy analysis is posed over a calibration set \(\{(\mathbf{x}_i,\mathbf{y}_i)\}_{i=1}^n\) through
$$
\mathbf{Y} = \mathbf{\Phi}\mathbf{W},
$$
followed by a column-subset selection objective,
$$
\min_{S\subset [N_e],\,|S|=m}\left\|\mathbf{Y}-\mathbf{\Phi}_{:,S}\mathbf{W}_{S,:}\right\|_F^2.
$$
The exact selection is framed as NP-hard, so the method ranks micro-experts by a decoding-time energy score. The paper first motivates
$$
\epsilon_{\mathrm{sup}} = \sum_{i\in S^C}\|\mathbf{\Phi}_{:,i}\|_2^2 \|\mathbf{w}_i\|_2^2,
$$
then defines
$$
\mathcal{E}_i = \left[(1-\alpha)\|\mathbf{\Phi}_{:,i}\|_2^2 + \alpha\|\mathbf{\Phi}_{:,i}\|_\infty^2\right]\cdot \|\mathbf{w}_i\|_2^2.
$$
Lower energy indicates greater redundancy.

CAMERA-P uses this ranking for **joint cross-matrix pruning**. For pruning ratio \(\lambda\), it removes the lowest-energy \(\lambda\) fraction of micro-experts in each MoE layer and deletes the corresponding row in \(\mathbf{W}^{\mathrm{up}}\), row in \(\mathbf{W}^{\mathrm{gate}}\), and matching column in \(\mathbf{W}^{\mathrm{down}}\). The layerwise update is written as
$$
\mathbf{W}^{\mathrm{up}} \leftarrow \mathbf{W}^{\mathrm{up}}_{S,:},\quad
\mathbf{W}^{\mathrm{gate}} \leftarrow \mathbf{W}^{\mathrm{gate}}_{S,:},\quad
\mathbf{W}^{\mathrm{down}} \leftarrow \mathbf{W}^{\mathrm{down}}_{:,S}.
$$
Because the pruning unit is structurally coupled across all three matrices, the resulting FFN is reduced in width without introducing unstructured sparsity.

The empirical evaluation centers on **Deepseek-MoE-16B**, **Qwen2-57B-A14B**, and **Qwen3-30B-A3B**, with calibration on **Wikitext2**, **128 sequences**, and **2048 tokens**. Reported pruning ratios are **20%**, **40%**, and **60%**. The method consistently exceeds **NAEE** and **\(D^2\)-MoE** on average zero-shot accuracy, with particularly strong behavior at aggressive pruning. On **Qwen2-57B-A14B** at **60% pruning**, the reported average scores are **51.40** for NAEE, **56.32** for \(D^2\)-MoE, and **65.17** for CAMERA-P. The paper also reports that complete micro-expert analysis of **Qwen2-57B-A14B** takes **less than 5 minutes** on a single **NVIDIA A100-40GB GPU**, and that pruning completes in about **0.1 GPU hours**. This suggests that CAMERA-P is designed not only for parameter reduction but for **actual structural speedup**, which the paper quantifies on **Deepseek-MoE-16B** as **\(1.06\times\)**, **\(1.42\times\)**, and **\(1.48\times\)** at **20%**, **40%**, and **60%** pruning.

## 3. CAMERA-P as a P-channel FDMA-CDMA CAOS camera mode

In the CAOS literature, CAMERA-P is not introduced as a separate formal acronym; rather, it is effectively the **CAOS camera operated in a “P-channel” FDMA-CDMA configuration** [2107.06175]. The underlying platform is the **CAOS (Coded Access Optical Sensor) camera**, a programmable “thinking camera” based on a **DMD spatial modulator** and **point detectors**. The hybrid mode combines three coding dimensions: **space coding**, **time coding** with Walsh CDMA, and **frequency coding** with FDMA carriers. The paper describes this as a **space-time-frequency triple coding design**.

The encoded photocurrent is modeled as
$$
i(w,t)=G \sum_{n=1}^{N}\sum_{m=1}^{M} I_{mn} d_{mn}(w,t),
$$
with
$$
d_{mn}(w,t)=C_{mn}(t) f_{mn}(w,t).
$$
Here \(I_{mn}\) is the irradiance of the \((m,n)\)-th CAOS pixel, \(C_{mn}(t)\) is the Walsh CDMA code, and \(f_{mn}(w,t)\) is the FDMA frequency assignment. The decoding chain performs **FFT analysis** during each CDMA bit duration \(T\), assembles the **\(P\) spectral peaks** over **\(W\)** bits, correlates them with the **\(J\)** Walsh code sequences, and maps the recovered irradiances back to the spatial grid. The key counting relation is
$$
J = \lceil Q/P \rceil,
$$
where \(Q\) is the total number of CAOS pixels.

The parameter \(P\) denotes the number of simultaneous FDMA channels, and the paper states that the demonstrated FDMA-CDMA mode operates **\(P\) times faster** than the equivalent linear HDR FM-CDMA mode. The example given is \(Q = 2035\), \(P = 8\), \(J = 255\), and Walsh length \(W = 256\) instead of **2048**, implying an **8-fold** reduction in encoding time. The stated benefits are **high security**, **linear HDR**, **high SNR**, and **higher speed**. The HDR claim is tied to FFT/DSP-based spectrum analysis, while the SNR claim is tied to simultaneous multi-pixel detection in the CDMA mode.

The active-mode demonstration uses **\(P=3\)** LEDs driven at **\(f_1 = 25\) kHz**, **\(f_2 = 29\) kHz**, and **\(f_3 = 35\) kHz**, each with distinct optical spectral content. The paper emphasizes that this permits simultaneous capture of **\(P\)** images **without time-multiplexed slots** and **without a tunable optical filter**. In that sense, CAMERA-P denotes a **parallel active sensing camera** whose rate and coding burden scale with the number of FDMA channels, rather than a separate camera body.

## 4. CAMERA-P as disturbance-free single-pixel imaging with \(P=0.5\)

In the supplied single-pixel-imaging mapping, CAMERA-P corresponds to the **single-pixel imaging camera based on complementary detection and optimized Bernoulli modulation**, called **CSPI** in the paper itself [2204.07711]. The architecture uses a **DMD** to project binary coded patterns, records **two complementary measurements simultaneously**, and subtracts the detector outputs. One detector sees \(A^i(x)\), the other sees \(1-A^i(x)\). The recorded intensities are
$$
Y_{\rm up}^i=\int A^i(x)\left(I_0T(x)+I_b^i\right)\,dx + I_{n\text{-}up}^i,
$$
$$
Y_{\rm down}^i=\int \left(1-A^i(x)\right)\left(I_0T(x)+I_b^i\right)\,dx + I_{n\text{-}down}^i,
$$
and the differential CSPI measurement is
$$
Y_{\rm CSPI}^i = Y_{\rm up}^i-Y_{\rm down}^i.
$$
This yields
$$
Y_{\rm CSPI}^i = I_0\int (2A^i(x)-1)\,T(x)\,dx + I_b^i\int (2A^i(x)-1)\,dx + I_{n\text{-}up}^i-I_{n\text{-}down}^i.
$$

The disturbance-cancellation mechanism depends on **Bernoulli-distributed binary patterns** with
$$
A^i(x)\in\{0,1\}, \qquad \Pr[A^i(x)=1]=P,\quad \Pr[A^i(x)=0]=1-P.
$$
When **\(P=0.5\)**, the mean is exactly \(\langle A(x)\rangle = P = 0.5\), and in the ideal balanced case
$$
\int (2A^i(x)-1)\,dx = 0.
$$
Under the paper’s stated assumption that the disturbance light is **spatially uniform on the DMD plane** but may vary arbitrarily from one measurement to the next, the term involving \(I_b^i\) vanishes. The paper therefore describes the camera as **disturbance-free** under those conditions. The reconstruction retains the standard correlation SPI framework:
$$
O_{\rm SPI}(x) = \frac{1}{K}\sum_{i=1}^{K} \left(A^i(x)-\langle A(x)\rangle\right)Y^i,
$$
$$
O_{\rm CSPI}(x) = \frac{1}{K}\sum_{i=1}^{K} \left(A^i(x)-\langle A(x)\rangle\right)Y_{\rm CSPI}^i.
$$

The paper compares standard SPI, differential SPI (DSPI), and CSPI. DSPI is given as
$$
O_{\rm DSPI}(x)=\frac{1}{K-1}\sum_{i=1}^{K-1} \left(A^{i+1}(x)-A^i(x)\right)Y_{\rm DSPI}^i,
$$
with \(Y_{\rm DSPI}^i = Y^{i+1}-Y^i\). The reported simulations show that **DSPI** can mitigate slowly varying background, but **fails when the disturbance fluctuates rapidly or strongly relative to the signal**, whereas **CSPI** remains effective even when disturbance intensity is rapidly varying, random, and much stronger than the target return. With sinusoidal disturbance and irradiation SNR \(\varepsilon = 10\) dB, the reported correlation coefficient \(\beta\) is **0.659** for DSPI and **\(\beta \approx 1\)** for CSPI. The paper also reports explicit tests for \(P=0.4\), \(0.45\), and \(0.5\), with best reconstruction at **\(P=0.5\)**. For a **\(64\times64\)** case using a Hadamard-coded arrangement of Bernoulli patterns, removal of the first all-ones pattern leaves **\(K=4095\)** measurements. A plausible implication is that, within the paper’s assumptions, CAMERA-P denotes a **pattern-statistics-based cancellation scheme** rather than ordinary background subtraction.

## 5. CAMERA-P as the pCam6060 astronomical photodetecting system

In astronomical detector development, CAMERA-P directly refers to the **pCam6060 photodetecting system** developed at the **Special Astrophysical Observatory of the Russian Academy of Sciences** and optimized for **photometric observations** [2412.10109]. The system is built around the **GSENSE6060BSI** back-illuminated CMOS detector with **\(6144 \times 6144\)** active pixels and **\(10\,\mu\text{m}\)** pixel size. The paper reports a **full-frame readout rate of 11 fps**, communication via a **fiber-optic line** at distances of **up to 50 m**, and real-time recording of video data to the computer hard drive.

The detector’s reported spectral response is **200–1040 nm**, with **minimum QE 10%**, **maximum QE 95% at 580 nm**, and **QE 58% at 850 nm**. The BSI architecture is emphasized because, unlike front-illuminated devices, it does not show the long-term residual bulk image effect characteristic of many FSI detectors. The paper does note a small residual lag after saturation: after reset, about **3 e\(^{-}\)** remain in saturated pixels in the next frame. This is one of the reasons the system is presented as particularly suitable for faint-object photometry and long-exposure observation methods.

A distinctive feature is the **simultaneous readout through two 12-bit video channels** with different gain settings and their controller-level combination into a single **16-bit frame**. The readout modes are **LG**, **HG**, and **HDR**. The paper describes the HDR method as using **multiplicative and additive coefficients** derived from measured transfer characteristics so that the combined response is linear and shows no **gain shift** or **dispersion shift** at the junction between channels. The reported gains are **\(23.8\ e^{-}/\text{ADU}\)** for LG, **\(2.39\ e^{-}/\text{ADU}\)** for HG, and **\(1.43\ e^{-}/\text{ADU}\)** for HDR. The corresponding readout noises are **23.6 e\(^{-}\) rms**, **3.47 e\(^{-}\) rms**, and **3.19 e\(^{-}\) rms**; the full well capacities are **93,200 e\(^{-}\)**, **9,300 e\(^{-}\)**, and **91,500 e\(^{-}\)**; and the dynamic ranges are **71.9 dB**, **68.6 dB**, and **89.1 dB**.

The photometric characterization further reports **non-linearity** of **0.63%** in LG, **0.80%** in HG, and **0.69%** in HDR; **photoresponse non-uniformity 0.5%**; **gain instability 0.064%**; **image lag 3–4 e\(^{-}\)/pixel**; and **dark current 0.2–0.3 e\(^{-}\)/s/pixel**. Thermal stability is critical because the gain dependence is reported as
$$
\Delta G \approx 0.32\% \text{ per } 1^\circ\text{C}.
$$
The controller therefore stabilizes the detector temperature to **\(\pm 0.1^\circ\)C**, which keeps gain instability below **0.064%**. Cooling uses **two-stage Peltier elements**, and the operating temperature can be brought to **\(60^\circ\)C below ambient radiator temperature**. The body dimensions are **190 × 190 × 170 mm**, and the system is described as **moisture-proof**. Within the astronomical context, CAMERA-P therefore designates a large-format, dual-gain, BSI-CMOS photometric camera emphasizing QE, readout-noise suppression, and HDR linearity.

## 6. CAMERA-P in high-energy astrophysics instrumentation

A separate usage associates CAMERA-P with fast Cherenkov-camera instrumentation for high-energy air showers. In the **PBR** proceedings contribution, the instrument is consistently described as the **Cherenkov Camera (CC)**, and the supplied mapping states that the query term appears to refer to this same package on board **POEMMA-Balloon with Radio (PBR)** [2511.15636]. The CC is a **2048-pixel SiPM camera** operating over **320–900 nm** with **10 ns integration time**. Its focal surface consists of **four rows of 8 SiPM arrays**, each array containing **\(8 \times 8\)** channels, yielding
$$
4 \times 8 \times 8 \times 8 = 2048
$$
pixels. The reported field of view is **\(12^\circ \times 6^\circ\)** with pixel angular scale of about **\(0.2^\circ\)** per pixel. The optical system is a **Schmidt-type telescope** with **1.1 m aperture**, approximately **1.6 m radius of curvature**, an **aspheric corrector plate** made of UV-transparent PMMA, and a **segmented primary mirror** of **12 vacuum-slumped borosilicate glass elements**. A distinctive element is the **bi-focal optical design**, implemented with a PMMA **1D prism array** mounted **100 mm in front of the focal plane**, which splits incoming light into two spatially separated spots and supports a coincidence-based trigger. The readout is based on the **MIZAR ASIC**, which handles **64 channels**, samples at **200 MHz**, provides **256 memory cells**, and offers a single-slope ADC with programmable resolution of **7–12 bits**. The trigger requires that **at least two pixels exceed the thresholds**.

A different but related paper uses CAMERA-P for the modular **SiPM camera and readout system** developed for the **Trinity Demonstrator** and the **EUSO-SPB2 Cherenkov Telescope** [2406.08274]. Here the camera is optimized for Earth-skimming **PeV–EeV tau neutrinos** observed with the imaging atmospheric Cherenkov technique. Two versions are reported: a **256-pixel** camera for Trinity using **Hamamatsu S14161-6050HS SiPMs**, and a **512-pixel** camera for EUSO-SPB2 using **Hamamatsu S14521-6050AN SiPMs**. The front end is built around the **eMUSIC ASIC**, and the signals are sampled and digitized with the **AGET** system at **100 MS/s** and **12-bit** resolution. Both cameras are **liquid-cooled**. Each camera module contains a **4×4 matrix of SiPMs**, front-end **Sensor Interface and Amplification Board (SIAB)** electronics, and, where required, adaptor boards for curved focal planes. The AGET chain uses a **512-cell buffer depth** and yields a **5.12 μs** trace length.

The quantitative characterization is unusually detailed. The selected operating point is where PDE reaches **90% of its maximum**, corresponding to about **9% relative overvoltage** for the reference device at room temperature, or about **44 V** bias. The paper reports breakdown-voltage temperature dependence of **37 mV/°C**, direct optical crosstalk of around **~1%**, afterpulsing of about **5%**, and a recovery time constant of about **100 ns**. Signal-chain calibration gives **\(10.07 \pm 0.09\)** digital counts per photoelectron, with linearity up to about **300 photoelectrons**. The raw pulses are slowed for AGET compatibility by a **third-order low-pass Butterworth filter** with **15 MHz** cutoff, producing rise time of about **20 ns** and output pulses with **~30 ns FWHM**. Flatfielding is performed through per-pixel bias trimming according to
$$
\Delta U_{\rm bias} = 4.5\,{\rm V}\cdot \frac{A_c - A_p}{A_c},
$$
and after flatfielding the reported pixel-response standard deviation is about **0.05**.

Taken together, these high-energy uses show that CAMERA-P can denote either a mission-specific Cherenkov focal-plane package or a modular SiPM readout architecture for fast air-shower imaging. The shared technical pattern is a focus on **fast timing**, **SiPM pixelation**, and **coincidence-capable electronics**, but the papers describe distinct camera systems rather than a single unified platform.

Source: https://www.emergentmind.com/topics/camera-p