A
Catalogue mass
The published typical weight for that class. It ignores the box entirely, and is therefore immune to a loose box, which is the common failure.
Technology
Detection.Scale.Then mass.
Three steps, each checkable on its own. The interesting problems were in the second and third.
01
Detection
SAAF draws a tight box around each item rather than a pixel mask. Litter in an aerial frame is small, about 60 pixels on its longest side in a 3840×2160 capture, so every frame is read at full resolution rather than downscaled.
Both scores below were measured on imagery the model had never seen. 100% is perfect.
99.2% - Of the litter was found and correctly namedA box counts as right when it covers at least half of the item. The everyday score, written mAP@50.
68.2% - Was boxed tightly enough to measure fromSame test, but the box has to fit almost exactly. Harder by design, and the number size and mass depend on. Written mAP@50-95.
What mAP is
One score for two questions: how much did it find, and how often was it wrong?
Mean average precision is the standard score for object detectors. It combines how much of the real litter was found with how few false alarms were raised, averaged over all six classes. The number after the @ is how tightly a box must fit the item before it counts: at 50 it must cover half the item, at 50-95 it is scored again and again up to a near-perfect fit and the results averaged. A loose box passes the first test and fails the second.
02
Why the box has to be tight.
A detection box is an axis-aligned rectangle. Turn a 22×7 cm bottle by 45° and the smallest upright box that contains it is a 20.5×20.5 cm square,2.7× the object's real area. A detector that is casual about orientation draws loose, square boxes, and every area and mass downstream inherits the error.
So box tightness is scored on its own, as mAP@50-95, and treated as the number that matters. Getting it right was the single largest improvement in the project, and it came from how the boxes are drawn rather than from a bigger model.
03
Scale: one number everything depends on.
Pixels become centimetres through the ground sample distance. Everything downstream is a multiple of this one number, so it is the only value that has to be right for the areas to be right. It needs the camera model and the flight altitude, both of which the pilot already knows.
GSD = sensor_width × altitude ÷ (focal_length × image_width)- sensor_width from the camera model
- altitude from the flight log
- focal_length from the camera model
- image_width 3840 px
04
Two masses per detection, because they fail differently.
B
Size-adjusted mass
Scales that weight by the item's longest side against a reference length. Length rather than area, because a loose box inflates area quadratically: a 2× area error would become a roughly 3× mass error under volumetric scaling. The ratio is clamped hard for the same reason.
How far unverified guesses were off
When estimates were checked against measured values, guesses missed by 43 to 59%. A glass bottle estimated at 350 g weighs about 220 g. Every weight in the table now has a published source.
05
Honest limits.
Stated here so nobody has to discover them in a report.
- 01
Mass is an estimate
It comes from reference weights, not a weighing, and it is reported as one everywhere it appears.
- 02
Altitude has to be right
Every area and mass scales linearly with the recorded altitude. A wrong altitude is a wrong survey.
- 03
Only the visible footprint counts
Partially buried or submerged items are under-measured. What the camera cannot see is not counted.
- 04
One flight is one moment
Waste moves with the tide and the season. A single survey is a snapshot; hotspots and trends need flights repeated across both.
- 05
Six classes
Anything outside them is not counted rather than misfiled. New classes can be added.
