Two containers for the same geometry
Every 3D sensor has to put its measurements somewhere, and the industry settled on two containers. A depth map is an image whose pixels store distance instead of color: a grid, viewed from exactly where the sensor sits. A point cloud is a bag of coordinates: thousands of (x, y, z) points floating in space, viewable from anywhere. Same scene, two representations, and most pipelines convert freely between them. Knowing which one a system uses tells you a lot about what it can and cannot do.
What is the difference between a depth map and a point cloud?
A depth map is organized and viewpoint-locked: every pixel has a defined neighbor, which makes image-style processing (filtering, segmentation, tracking) fast and simple, but the data only makes sense from the sensor's own perspective. A point cloud is unorganized and viewpoint-free: points can be merged from multiple sensors, rotated, and measured against each other in real-world units, at the cost of heavier processing. Depth maps suit a single fixed sensor; point clouds suit mapping, scanning, and multi-sensor scenes.
Where each representation comes from
A Time of Flight camera natively produces a depth map: its detector is a pixel grid, and each pixel reports a distance (how that measurement works). Converting to a point cloud is straight trigonometry: knowing each pixel's viewing angle (the camera's intrinsics), every depth value becomes an (x, y, z) point. Scanning LiDAR, by contrast, produces points natively as its beam sweeps, and building an image-like structure from them takes resampling; that architectural difference is part of ToF vs LiDAR.
Why counting pipelines stay in the depth map
An overhead people counter watches one fixed scene from one fixed position, which is exactly the geometry a depth map is built for. A person below appears as a connected region of near pixels; tracking connected regions frame to frame is fast, reliable, and cheap enough to run on the sensor itself. Nothing about the task rewards converting to a point cloud, so production counting firmware rarely bothers. The result reaches the platform as counts and trajectories, not as raw geometry in either format. The category view is in the ToF people counter guide; the deeper 3D-sensing story is in 3D people counting, explained.
Neither one is a photograph
A depth frame of a doorway shows human-shaped regions of distance. A point cloud of the same doorway shows the same shapes as floating dots. Neither contains a face, clothing color, or anything that identifies a person, and no post-processing can conjure identity out of geometry that never captured it. That property is structural, not a setting, which is why depth sensing keeps winning the measurement jobs where cameras create more legal surface than value. The wider argument lives in people counting without cameras.
---