← BACK TO WORK

05 / CASE STUDY

MACHINE LEARNING / COMPUTER VISION · 2025

ADVERSEWEATHER

A local computer-vision experiment measuring vehicle detection under rain, fog, and snow with BDD100K and YOLOv5 transfer learning.

DATASETBDD100K
MODELYOLOv5
VIEW REPOSITORY ↗
WEATHER / DETECTION / EVALUATION

01 / THESIS

VISIBILITY
CHANGES
THE RESULT.

Adverse Weather AV Detection treats visibility as an evaluation condition: the same object-detection task is examined across rain, fog, and snow rather than presented as a production driving system.

The project keeps the result measurable and bounded. BDD100K supplies the dataset context, YOLOv5 supplies the detector, and recorded evaluation shows where performance becomes harder to interpret.

02 / THE EXPERIMENT

THE INPUT BECOMES HARDER TO SEE.

BDD100K frames move through weather conditions, YOLOv5 detection, and recorded evaluation without implying a production system.

01BDD100K FRAMEinput image
02WEATHER FIELDrain · fog · snow
03YOLOv5car detection
04EVALUATIONmeasured result

03 / WEATHER FIELD

RAIN, FOG, AND SNOW REMAIN DISTINCT CONDITIONS.

01 / CONDITIONRAINBDD100K / EVALUATION SET
02 / CONDITIONFOGBDD100K / EVALUATION SET
03 / CONDITIONSNOWBDD100K / EVALUATION SET

04 / EVALUATION

CAR CLASS / RECORDED EVALUATIONYOLOv5
PRECISION0.7516
RECALL0.5036
mAP500.5800
mAP50-950.3292

MODEL LIMITS BECOME MEASURABLE.

Recorded car-class results keep the experiment legible: the numbers describe evaluation performance, not deployment readiness.

05 / DECISIONS

MEASURE THE DIFFICULT INPUT.

01

Weather as evaluation domain

Rain, fog, and snow remain explicit conditions so degraded visibility is part of the experiment rather than hidden behind one aggregate result.

CONDITION / DETECTION / EVALUATION
02

Detection before interpretation

YOLOv5 is evaluated on the car class before any broader claim about autonomous driving or operational deployment is made.

CONDITION / DETECTION / EVALUATION
03

Measurement over anecdote

Precision, recall, mAP50, and mAP50-95 keep the experiment grounded in recorded model evaluation and visible limitations.

CONDITION / DETECTION / EVALUATION

06 / SIGNATURE

VISIBILITY INTO UNCERTAINTY.

OBSERVATIONCLEAR FRAMEINPUT IMAGE
VISIBILITYDEGRADED FIELDRAIN / FOG / SNOW
MEASURED LIMITEVALUATIONPRECISION / RECALL / mAP

07 / DEMONSTRATES

ADVERSE CONDITIONS, OBJECT DETECTION, MEASURED PERFORMANCE.

BDD100K CONDITIONS

YOLOv5 DETECTION

RECORDED EVALUATION

PROJECT / ADVERSE WEATHERMACHINE LEARNING / COMPUTER VISION2025 / LOCAL EXPERIMENT