M.TECH · DISSERTATION · FINAL VIVA BITS PILANI WILP · 2025–2026

Final Presentation ● May 2026

Road Quality Assessment
Using Computer Vision
and GPS-Based Mapping

Crowdsourced, dashcam-based road quality assessment with quality-aware route planning across 808 OSM segments in Kerala.

Student

Anish A.2024TM93051 · M.Tech Software Engg.

Supervisor

Dr. Kavya ManoharAdalat AI

Additional Examiner

Mr. Ashik SalahudeenAuxmoney GmbH

01 / 10

02 · The Problem Road Quality Assessment

02 ● The Problem

Navigation today optimises distance and time —
not the quality of the road you actually drive on.

01 · The Gap

Maps ignore surface quality.

Every major navigation system today optimises distance and ETA. Road comfort, vehicle wear, and surface safety are invisible.

02 · The Indian Context

Western models do not transfer.

High visual clutter, monsoon damage, varied surfaces (asphalt, concrete, gravel, dirt) — pothole detectors trained on Japan / US / EU datasets fail on Indian roads.

03 · The Data Gap

Municipal surveys are manual & infrequent.

Surveys happen once or twice a year, are not published openly, and cannot inform driver-facing routing decisions in real time.

Research
Question Can dashcam footage and computer vision automatically classify road quality at segment level, map it to a navigable graph, and route drivers around the worst of it?

02 / 10 Anish A. · 2024TM93051

03 · System Pipeline Six stages · dashcam → routing

03 ● End-to-End Pipeline

From dashcam video to quality-aware routes.

Dashcam
Ingestion

1080p footage from Kerala roads; FFmpeg frame extraction @ 1 fps.

~34k frames
per camera

›

OCR + GPS
Filtering

Tesseract OCR on GPS overlays; static / night / out-of-bounds frames dropped.

~26k frames
annotatable

›

Consensus
Annotation

Gamified platform · 2-annotator agreement with 3rd-annotator tiebreaker.

3,216 imgs
53 annotators

›

Model
Training

Architecture sweep + 72-trial HPO. Swin-Small fine-tuned at lr=3e-5.

15 archs
72 HPO trials

›

GPS
Snapping

osmnx + UTM projection; nearest-edge within 35 m; per-edge aggregation.

808 OSM
segments

›

Routing &
Map Server

Flask + Leaflet UI; shortest vs. smoothest path via quality-penalised Dijkstra.

REST API
+ live map

3,216

Consensus images

808

OSM segments mapped

79.6%

Validation acc · 3-class

75%

Macro F1 · 3-class

03 / 10 Anish A. · 2024TM93051

04 · Dataset & Annotation 3,216 consensus-validated images

04 ● Dataset & Annotation

A crowdsourced dataset built with consensus and active learning.

Class Distribution · 3-class (production)

Good73.9%2,376

Bad15.9%511

Invalid10.2%329

5-Class Breakdown (pre-merge)

33.2%

40.6%

12.4%

Excellent 1,069 Good 1,307 Fair 399 Poor 112 Invalid 329

Active learning moved class ratio from 9.2:1 → 4.6:1 (Good:Bad).

Annotation Protocol

Two annotators → agreement, or third tiebreaker.

Gamified platform with leaderboard and consensus-accuracy feedback kept contributors engaged across multiple weeks.

Contributors

3,216

Confirmed labels

Dashcam sources

Active Learning Loop

Frames where max(softmax) < 0.7 are pushed up the queue.

Targets minority & ambiguous samples. The Bad class grew from 4.1% → 15.9% across iterations — directly addressing imbalance. Group-aware train/val split on video identity prevents near-duplicate leakage.

04 / 10 Anish A. · 2024TM93051

05 · Architecture Sweep 13 successful · Swin-Small wins

05 ● Model Selection

A 13-architecture sweep, ranked by macro F1.

5-class · fixed lr=1e-4 · weighted loss · ~1.6 h

#	Architecture	Params	Acc	Macro F1
01	swin_small ✓	50M	53.8%	47.0%
02	swin_tiny	28M	51.5%	46.1%
03	resnet34	21M	51.1%	45.4%
04	convnext_tiny	28M	50.4%	44.8%
05	resnet18	11M	49.7%	44.4%
06	vit_small	22M	49.8%	44.0%
07	efficientnet_b1	8M	49.4%	41.4%
8–13	efficientnet_b0/b2 · resnet50 · convnext_s/base · vit_base (efficientnetv2_s/m failed)	5–89M	45–50%	39–42%

Why Swin Wins

Hierarchical, shifted-window attention fits road texture.

Local 7×7 windows match pothole-scale features; shifted windows preserve cross-boundary context; multi-scale stages capture both crack-pattern texture and segment-level structure.

Why ViT Loses · Inverse Scaling

Capacity-data mismatch 97% train · 59% val

Isotropic transformers memorise rather than learn on a 3k-image domain dataset. The same pattern repeats across families: ConvNeXt-Base (89M) ranked 13th vs. ConvNeXt-Tiny (28M) at 4th; ViT-Base (86M) at 11th vs. ViT-Small at 6th. Data-limited, not capacity-limited.

05 / 10 Anish A. · 2024TM93051

06 · Hyperparameter Tuning 72 trials · ~14 hours

06 ● Hyperparameter Tuning

72 trials surfaced a counter-intuitive recipe.

Learning Rate

3e-551.9%

1e-451.7%

3e-451.7%

Backbone Freeze

0 epochs52.1%

5 epochs51.7%

10 epochs51.4%

Label Smoothing

none51.9%

0.151.6%

——

Class Weighting

unweighted52.0%

weighted51.5%

——

Final Production Configuration

model	swin_small
learning rate · scheduler	3e-5 · CosineAnnealingLR
backbone freeze	none — full fine-tune
loss · smoothing	unweighted CE · 0.0
optimiser · batch · stop	Adam · 32 · val-loss patience 30

Surprise · 01 · Backbone freezing hurts

ImageNet → Indian dashcam is a wide domain gap. Early layers need to adapt. Two-phase training underperformed end-to-end fine-tuning at every duration.

Surprise · 02 · Smoothing & weights add nothing

Swin already learns reasonably balanced representations at this scale. The bottleneck is the class boundaries themselves, not optimisation.

06 / 10 Anish A. · 2024TM93051

07 · Class Consolidation Data-driven 5 → 3

07 ● The Class Consolidation Decision

Five classes were the bottleneck — not the model.

Confusion matrices across all 13 successful architectures showed persistent confusion at the Excellent ↔ Good and Fair ↔ Poor boundaries. Annotator-disagreement data corroborated this — humans split on those exact pairs. The ambiguity was real, not a modelling failure.

5-class · Excellent / Good / Fair / Poor / Invalid

53.0% acc

57.0% macro F1

Production swin_small on the 5-class consensus dataset. Macro F1 was respectable but minority-class recall remained poor.

›

3-class · Good / Bad / Invalid · ADOPTED

79.6% acc

75.0% macro F1

Excellent + Good → Good; Fair + Poor → Bad. Directly answers the navigation question: is this segment safe to drive?

Insight The biggest single performance gain came from listening to the data — not from a deeper network or longer training. 53.0% → 79.6% acc · 57.0% → 75.0% macro F1.

07 / 10 Anish A. · 2024TM93051

08 · Final Model Performance swin_small · 3-class · 941 val samples

08 ● Final Model Performance

Strong on Good roads — conservative on Bad.

Confusion Matrix · 941 validation samples · Invalid excluded

Pred · Bad

Pred · Good

True
Bad

182

true negative · 53%

160

false positive · 47%

True
Good

false negative · 6%

561

true positive · 94%

Pessimistic edge aggregation at the segment level partially mitigates the 47% Bad miss-rate — a segment is only labelled Good if all observations agree.

Per-Class Classification Report

Class	Precision	Recall	F1	Support
Bad	83%	53%	65%	342
Good	78%	94%	85%	599
Macro avg	80%	73%	75%	941

79.6%

Overall accuracy

75%

Macro F1

808

OSM segments mapped

Known Limitation

53% recall on Bad — the model is conservative.

Residual imbalance & visual ambiguity at driving speed. Mitigated at segment level via pessimistic edge aggregation in routing.

08 / 10 Anish A. · 2024TM93051

09 · Live Demonstration Four-part walkthrough

09 ● Live Demonstration

From annotation through to navigation — end to end.

Annotation Platform

flask · jinja · 53 contributors · 3,216 confirmed labels

Road Quality Maps

folium · osmnx · 808 segments · 35 m snapping

Quality-Aware Route Planning

dijkstra · pessimistic / majority modes · unrated = 2.5×1.2

Navigation Server

flask · leaflet.js · REST + GeoJSON · port 5050

Project Home roads.anishsheela.com ↗

Reports · source code · live demos · maps · slides · dataset notes

09 / 10 Anish A. · 2024TM93051

10 · Conclusion & Future Work Thank you

10 ● Conclusion & Future Work

A complete, deployable pipeline —
dashcam to driver, in one system.

Project Home

roads.anishsheela.com

Reports · source · live demos · maps · slides ↗

Contributions

Empirical case for classification over detection.

17 YOLO variants benchmarked; best mAP50 of 20.3%.

Consensus-annotation methodology.

3,216 labels across 53 contributors · gamified + active learning.

Data-driven class consolidation.

5 → 3 merge drove val accuracy 53% → 79.6%.

End-to-end geospatial integration.

UTM snapping, dual aggregation, Flask + Leaflet nav server.

Future Work

01 · Open Release

Anonymise & release dataset and models

PII-stripped frames + trained weights for the research community.

02 · Deployment

Mobile navigation application

Phone-first routing app for in-field driver use.

03 · Coverage

Expanded fleet & beyond Kerala

More cameras, post-monsoon seasons, other Indian states.