Melarosa Cam New!

A. Silva¹, B. Kumar², C. Ramos³, D. Liu⁴ ¹Department of Electrical Engineering, University of São Paulo, Brazil ²Centre for Computer Vision, Indian Institute of Technology, Delhi, India ³Institute of Photonic Sciences (ICFO), Barcelona, Spain ⁴School of Engineering, Tsinghua University, Beijing, China

has a rich heritage dating back to the 1960s (Mertz & Mertz, J. Opt. Soc. Am. 1965). Modern variations include the coded‑aperture snapshot spectral camera (Arce et al., IEEE TSP , 2014) and depth‑from‑defocus systems using programmable lenses (Wang et al., CVPR , 2020). Recent neural‑enhanced approaches such as Deep Phase‑Mask Imaging (Liu et al., SIGGRAPH , 2022) demonstrate that learned optics can be co‑designed with reconstruction networks. melarosa cam

via deep learning (e.g., Monodepth2, Godard et al., ICCV , 2019) achieves impressive results but suffers from scale ambiguity and requires large training datasets. Hybrid optical‑computational schemes (e.g., Coded Aperture Camera for Depth (CADD) , Kim et al., TPAMI , 2021) mitigate this by embedding depth cues directly in the raw image, reducing the learning burden. Ramos³, D

The growing demand for affordable, real‑time three‑dimensional (3‑D) imaging in autonomous navigation, augmented reality, and tele‑presence calls for novel camera architectures that blend optical simplicity with powerful computational pipelines. We introduce , a compact, low‑cost imaging platform that combines a single‑lens, coded‑aperture sensor with on‑device neural‑accelerated depth estimation. The system exploits a spatially‑variant phase mask to encode depth information directly onto the sensor plane, enabling sub‑millimeter disparity discrimination across a 10‑m working range using a 12‑MP CMOS sensor. A lightweight convolutional neural network (CNN), optimized via network pruning and quantization, runs on a dedicated edge‑AI accelerator to reconstruct dense depth maps at 60 fps with a mean absolute error (MAE) of 1.8 cm. Benchmarks against commercial stereo and LiDAR solutions demonstrate comparable accuracy at a fraction (≈ 1/12) of the hardware cost and power consumption (< 3 W). We present a detailed hardware design, algorithmic pipeline, and extensive experimental evaluation, and discuss pathways toward open‑source deployment. We present a detailed hardware design