Stage 1 finds objects — what YOLO is good at. Stage 2 verifies the material of
each crop — what a dedicated classifier is good at — and filters false
alarms.
S1
Upload image
A JPG, PNG, or WEBP photo of waste. Saved scans stay in browser history for review and hard-case collection.
S2
Stage 1 — YOLO26n localization
YOLO26n proposes a box around every candidate object. 100-epoch retrain, later hard-negative fine-tuned for field small-object recall: 70.9% mAP50, 77.8% precision on clean validation.
S3
Crop extraction & size filter
Each kept box is cropped with 10 px padding; boxes smaller than 24×24 px are rejected before classification.
S4
Stage 2 — crop verification
A ConvNeXt-Tiny + 637-handcrafted-feature classifier re-checks every crop across 7 classes, including a Background veto. Fine-tuned on the detector's own crops so training matches production.
S5
Score fusion & context prior
Per-box score fuses detector and classifier confidences, then a scene-context prior adjusts the class scores.
S6
Decision gate & bin routing
Class thresholds and the waste-state gate pick the final decision. Uncertain detections route to review instead of a bin — conservative by design.
Every number below is tied to a report in the repository.
The evaluation sets were audited with perceptual hashing; duplicated eval images were
quarantined before these metrics were measured.
Classifier accuracy, clean test93.77%ConvNeXt + 637-feature classifier, 2,151 quarantined-test imagesDetector mAP50, clean val70.9%YOLO26n, hard-negative fine-tuned, quarantined splitEval leakage found & removed2,40643% of the old classifier test setCross-domain accuracy39.8%Studio-trained features on real-world images
Evaluation auditcompleted 2026-07-02
We found our own evaluation bug and corrected the numbers.
The merged classifier dataset ingests the same photos through different sources.
Perceptual hashing found 2,406 leaked test images (43%) in an earlier EfficientNetB0
classifier's evaluation set, correcting its accuracy from an inflated 94.30% to an
honest 91.77%. The same audit method, applied to the deployed ConvNeXt's dataset,
found a smaller leak and is the basis for the 93.77% figure above.
runs/audits/MODEL_RISK_AUDIT.mdaudit trail
Deployment truthlive model active
The API serves the real classifier and YOLO26n localizer.
The classifier is fine-tuned on the detector's own crops: accuracy on real
detector output rose from 76.9% to 88.9% without losing clean-crop accuracy.
When the API is offline the page says so instead of pretending.
Detector precision77.8%
Clean validation split. Tunable via confidence threshold.
Detector recall62.6%
Down slightly from 65.0% studio recall after a hard-negative fine-tune that traded it for +16 points of field small-object recall (see Limits).
Detector mAP50-9550.0%
Hard-negative mining checkpoint, promoted to the live localizer.
Classical ML baselineexplainable evidence
73.8%
ExtraTrees on 637 handcrafted features per crop (8 spatial, 9 FFT, 44 color, 576 HOG). Kept for comparison and explainability evidence beside the deep pipeline.
Saved scans
Recent scans
Restore a saved scan to bring back the image, result, confidence, route, and boxes.
History lives in your browser.
0 saved
Honesty section
Measured limits, and what was done about them.
Domain shiftPartly fixed, partly open
Fixed: the classifier was fine-tuned on the detector's own crops, lifting
production-distribution accuracy from 76.9% to 88.9%. The detector was re-tuned
for the field domain too: lowering the confidence threshold and hard-negative
mining on new field data (PlastOPol) lifted held-out field small-object recall
from 38.6% to 54.7%, at a small cost to studio mAP50 (72.1% → 70.9%). Open:
on fully unseen capture sessions the detector still drops to 0.444 mAP50; the
gap is narrowing with new data, not fully closed.
Tiny objectsResolution ruled out; new data confirmed as the lever
A 960px training run was evaluated: validation unchanged, unseen-batch recall
up 3.9 points — not enough to chase. A later hard-negative mining fine-tune
on new field data confirmed the real lever: small-object recall on that held-out
domain rose 38.6% to 54.7% from new data alone, no resolution or architecture
change. Small objects remain the main source of misses; training-data diversity
is the binding constraint.
Waste-state decisionConservative by design
The waste / not-waste / review decision applies conservative rules on top of the
verified material class; uncertain detections route to review instead of a bin.
Method note — metrics on this site come from the leakage-audited evaluation sets. The audit scripts, quarantine manifest, and reports live in the project repository.
A Final Year Project for automated waste understanding. A two-stage deep-learning
pipeline detects and classifies waste in real photos, backed by a classical
machine-learning branch for explainable evidence. Every headline metric survives a
leakage-audited evaluation set.
Live deployment
The real model, served publicly
The trained YOLO26n localizer and ConvNeXt classifier run behind a public API on Hugging Face Spaces.
