A minimalist vision system uses the smallest number of pixels needed to solve a vision task. While a traditional camera uses a large grid of square pixels, a minimalist camera uses freeform pixels that can take on arbitrary shapes to increase their information content. The hardware of a minimalist camera can be modeled as the first layer of a neural network, where the subsequent layers are used for inference. Training the network for any given task yields the shapes of the camera's freeform pixels, each of which is implemented using a photodetector and an optical mask. We have designed minimalist cameras for monitoring indoor spaces (with 8 pixels), measuring room lighting (with 8 pixels), and estimating traffic flow (with 8 pixels).

Modern cameras capture images with millions of square pixels arranged in a dense grid. This design philosophy assumes that high-resolution images are always necessary for visual understanding. However, for many practical vision tasks, most of the captured information is redundant. A surveillance camera monitoring whether a room is occupied does not need megapixel resolution. A sensor measuring ambient lighting does not need to capture a detailed image of the room. This observation motivates us to ask: what is the minimum number of measurements needed to solve a given vision task?

We introduce the concept of freeform pixels — pixels that are not constrained to be small squares arranged in a grid, but can instead take on arbitrary spatial shapes. By allowing each pixel to integrate light over a carefully designed region of the scene, a single freeform pixel can capture far more task-relevant information than a single square pixel. The key insight is that the shape of a freeform pixel acts as a spatial filter — it determines what combination of scene content contributes to the pixel's measurement. By optimizing these shapes for a specific task, we can design cameras that achieve high performance with an extremely small number of pixels.

We model the entire minimalist vision system as a neural network. The first layer of the network represents the camera hardware: each neuron in this layer corresponds to one freeform pixel, and its weights encode the pixel's spatial sensitivity pattern — effectively, the shape of its optical mask. The subsequent layers form a standard inference network that maps the pixel measurements to the desired output (e.g., room occupancy count, lighting parameters, or traffic density). During training, both the pixel shapes and the inference network weights are optimized jointly using backpropagation, ensuring that the hardware and software are co-designed for optimal task performance.

Each freeform pixel is physically realized using a photodetector paired with an optical mask. The mask is a binary or grayscale transparency pattern placed in the image plane that determines which parts of the scene contribute to the photodetector's measurement. The mask pattern is derived directly from the learned weights of the corresponding neuron in the first layer. For binary masks, we apply a thresholding operation during fabrication; for grayscale masks, halftone printing techniques are used. This hardware implementation is low-cost and compact, requiring only standard photodetectors and printed masks.

The shape optimization of freeform pixels can be understood through the lens of information theory. A square pixel in a traditional camera captures a uniform spatial average over a small, fixed region. In contrast, a freeform pixel captures a weighted spatial average over a potentially large, task-optimized region. The information content of a freeform pixel is therefore determined not just by signal-to-noise ratio, but also by how well its spatial sensitivity pattern aligns with the task-relevant features of the scene. Our optimization procedure naturally discovers pixel shapes that act as matched filters for the most informative spatial patterns.

An important property of our approach is that it naturally tends to preserve the privacy of individuals in the scene. Because the minimalist camera captures only 8 highly abstracted measurements rather than a detailed image, it is fundamentally incapable of recording identifiable visual information about people. This is a significant advantage for applications such as smart buildings, occupancy counting, and ambient intelligence, where privacy concerns have historically been a major barrier to adoption. The privacy preservation is inherent to the hardware design rather than relying on software-based anonymization, making it robust against adversarial attacks.

We evaluate the minimalist camera on three vision tasks: indoor occupancy monitoring, room lighting estimation, and traffic flow measurement. For each task, we compare our approach against baselines including conventional cameras with downsampled images (to match the number of pixels) and random pixel configurations. In the occupancy monitoring task, our 8-pixel minimalist camera achieves 97.5% accuracy in distinguishing between 0-5 occupants, compared to 82.3% for an 8-pixel downsampled conventional image and 68.1% for random pixel masks. This demonstrates the critical importance of task-optimized pixel shapes versus naive spatial sampling.

For the lighting estimation task, the minimalist camera must estimate the spatial distribution of light sources and their intensities in a room from only 8 measurements. Our optimized freeform pixels achieve a mean angular error of 4.2 degrees for light direction and 8.7% relative error for intensity, significantly outperforming downsampled images (11.8 degrees, 19.3%). The learned pixel shapes reveal that the system has discovered directional sensitivity patterns that act as spatial probes for different regions of the room. For traffic flow estimation, the minimalist camera achieves a mean absolute error of 2.1 vehicles per minute from only 8 freeform pixels on a busy highway, with self-powered operation enabled by the minimal sensor requirements.

We have presented minimalist vision — a new paradigm for visual sensing that uses the fewest possible pixels to perform a task. By introducing freeform pixels with task-optimized arbitrary shapes and modeling the camera as the first layer of a neural network, we enable end-to-end co-design of hardware and software. Our results on three practical applications demonstrate that 8 freeform pixels can match or approach the performance of conventional cameras with orders of magnitude more pixels. Beyond performance, our approach offers inherent advantages in privacy preservation, energy efficiency, and form factor. We believe minimalist vision opens a new direction in computational imaging, where the goal is not to capture as much visual information as possible, but to capture precisely the information needed for the task at hand.