Skip to main content

Module vision

Module vision 

Source
Expand description

std/vision/v1 — perception outputs (§24.3).

TypeLayout
Detections{boxes: List<FixedSizeList<Float32, 4>>, scores: List<Float32>, labels: List<UInt32>}
Keypoints{points: List<FixedSizeList<Float32, 2>>, scores: List<Float32>, labels: List<UInt32>}
Mask{width: UInt32, height: UInt32, data: List<UInt16>}

The three List columns of Detections and Keypoints are index-aligned: entry i of scores is the confidence of entry i of boxes. Nothing in the columnar layout enforces that — three lists of different lengths are a well-formed Struct — so both types check it on encode and on decode, which is the difference between “the payload is malformed” and “the labels are silently off by one”.

§Examples

use astrs_node_api::message::{AstrsMessage, BoundingBox, Detections};

let detections = Detections::new(
    vec![BoundingBox::new(0.0, 0.0, 10.0, 20.0)],
    vec![0.93],
    vec![7],
)?;
let batch = detections.to_record_batch()?;
assert_eq!(Detections::from_record_batch(&batch)?.len(), 1);

Structs§

BoundingBox
One axis-aligned detection box, in image pixel coordinates.
Detections
Detection boxes with per-box confidence and label — std/vision/v1/Detections.
Keypoint
One landmark, in image pixel coordinates.
Keypoints
Landmark sets with per-point confidence — std/vision/v1/Keypoints.
Mask
A per-pixel label plane — std/vision/v1/Mask, row-major.

Functions§

detection_boxes
The boxes column of a Detections payload, without decoding the rest.
mask_labels
The raw data column of a Mask payload, without copying the dimensions.