Is each stixel a detected object?

No. Objects may span many stixels, and grouping or recognition is usually an additional step.

Are stixels always one segment per column?

No. Multi-layer representations can include several segments within a column.

Do stixels contain only depth?

Not necessarily. Depending on the method, they may also carry semantic classes or other attributes.

AI Glossary / Computer vision and multimodal AI / Intermediate

Stixel

Definition

A stixel is a compact, vertically organized element used to represent parts of a scene in a structured visual-perception model. In the Stixel World family of methods, narrow image columns or segments summarize geometry and sometimes semantic information, providing an intermediate representation between individual pixels and complete recognized objects.

What it means

Dense images and depth maps contain large amounts of detailed information. A perception system may need a simpler representation that preserves useful scene structure without processing every pixel independently at every later stage.

Stixels group information into narrow vertical elements with properties such as image extent, depth, or class. Early formulations emphasized road scenes with free space and approximately upright obstacles; later formulations extend the representation and how it is estimated.

A stixel is not necessarily an object. Several stixels may describe one vehicle or building, while a single coarse element can cover geometry that requires further interpretation. The representation organizes evidence rather than automatically supplying complete object identity or motion understanding.

How it works

Obtain image, depth, disparity, or learned visual evidence according to the method. Divide the image into narrow columns and infer segments within them using a model of likely scene structure.

Assign each segment a compact set of attributes. Depending on the formulation, these can describe support surfaces, obstacles, sky, depth, semantic category, or other useful properties. Multi-layer versions can represent several vertically separated segments in a column.

Optimize a balance between agreement with the measurements and structural or complexity preferences. The result reduces a large pixel representation to fewer scene elements, but the compression necessarily discards some detail.

Use these elements for later tasks such as grouping, tracking, or free-space reasoning. Validate the assumptions against real scenes, including unusual slopes, overhanging structures, thin objects, missing depth, and occlusion. Different stixel systems should be evaluated according to their actual input and model rather than the name alone.

Three practical examples

1. Road-scene research: A perception pipeline summarizes visible obstacle structure and support surfaces in a compact representation for later processing.

2. Robotics visualization: Developers inspect vertically organized depth elements to understand what a scene model retained or discarded.

3. Object grouping: A downstream model combines neighboring stixels with compatible geometry or semantics into larger candidate objects.

Think of it this way

Think of sketching a street view using narrow vertical strips instead of drawing every pixel. The strips preserve broad positions and heights while omitting much of the texture.

The analogy is limited because real stixels can carry depth and other estimated attributes, not merely colored rectangles. Their boundaries and meanings are inferred under a model, and the compact sketch should not be mistaken for a complete geometric reconstruction.

Worked example

Consider a hypothetical image with width 1200 pixels. If a representation uses columns six pixels wide, it has 1200/6 = 200 columns.

If each column contains an average of three stixel segments, the scene is summarized by approximately 600 elements. A 1200-by-800 image contains 960,000 pixels, so there are 1600 input pixels per output element on average in this illustrative count.

This ratio is not a compression-byte ratio or a speedup claim. Each stixel stores multiple attributes, the algorithm needs intermediate data, and computation depends on the inference procedure.

If a thin pole occupies only two pixels horizontally, a six-pixel column representation may have difficulty preserving it distinctly, depending on the evidence and model. Choosing wider columns can reduce element count while increasing the risk of losing narrow structures.

Illustrative case study

Fictional case: A robotics research team uses a stixel representation to summarize depth images of a loading area. Large upright objects appear clearly, but thin support rods and overhanging shelving expose limitations in the initial model.

The team evaluates element width and a multi-layer formulation, then checks which structures remain visible in the representation. It keeps original depth data available for debugging and for tasks requiring finer detail.

The team does not infer operational safety from a neat visualization. It measures representation errors and downstream behavior on difficult scenes, treating the stixel model as one component of a larger perception system.

Common misconceptions and mistakes

Do not equate a stixel with a pixel, a voxel, or a complete object. It is a structured intermediate element with method-dependent attributes.

Do not assume all stixel methods require stereo cameras. Later approaches can use other evidence, including learned monocular predictions, with different uncertainty and scale considerations.

Do not assume the earliest road-scene assumptions fit every environment. Overhangs, slopes, and non-vertical structures can require richer models.

Do not report element-count reduction as an automatic accuracy or runtime improvement. The representation and the algorithm that constructs it have separate costs and limitations.

Business use

Stixels can help organize visual scene information for efficient downstream processing and interpretation. They are most relevant when the application's geometry fits the representation well enough that reduced detail is an acceptable tradeoff.

Evaluate retained structure, missed obstacles or surfaces, semantic errors where applicable, and processing latency. Compare the full pipeline with alternatives on representative conditions. A compact model may simplify later computation, but any loss of small or unusual structures must be understood in relation to the decisions the system supports.

Check your understanding

A representation reduces an image to 500 stixels. Does that mean it has detected exactly 500 objects?

Answer: No. The elements summarize scene structure. The number of physical objects requires separate grouping and interpretation.

Related terms

Disparity map
Stixel World
scene representation
free-space estimation
semantic segmentation
object grouping
stereo vision

Learn next

Draw a simple scene using fixed-width vertical strips and compare the result with the original outlines. Add an overhang and a thin pole to see why segment count, column width, and structural assumptions affect what can be represented.

Sources

Cordts and colleagues, The Stixel World: A Medium-Level Representation of Traffic Scenes, primary paper abstract: https://arxiv.org/abs/1704.00280 . The abstract was opened for the representation's scope. Earlier work by Badino, Franke, and Pfeiffer introduced the Stixel World. Counts and the loading-area case here are original illustrations.

Previous: STIT logic Next: Stochastic algorithm