Backyard Sweep

Turn the weekly "check for standing water" chore into a short, guided walk.

Live demo: https://mosquito-breeding-detector.onrender.com (Free server: the first load after a quiet spell can take up to a minute.)

Photograph anything around your home that could hold water. A YOLOv8 detector flags the containers that commonly breed mosquitoes, and each one comes with a plain-language tip for what to do about it.

Built for the Hacktoberfest 2026 Open-Source AI Challenge, Week 1: Touch Grass.

Example result on a photo from the VisText-Mosquito test split (not one of my own photos). Photo: Faiyaz and Jisan, CC BY 4.0.

Contents

1. Why this exists
2. Features
3. How it works
4. Model and results
5. API
6. Why open source mattered
7. Run it locally
8. Deploy on Render
9. Project structure
10. Limits and responsible use
11. Ideas for next steps
12. Credits and licence

Why this exists

Dengue is spread by Aedes mosquitoes, which breed in small containers of standing water: tires, vases, bottles, coconut shells, blocked drains. Health agencies advise emptying, scrubbing or covering these regularly. The hard part is not knowing what to do. It is noticing the containers you have stopped seeing.

Backyard Sweep turns that into a walk. You go outside, photograph what could hold water, and get a short to-do list with a tip for each spot.

Features

• Camera first. On a phone, "Take a photo" opens the camera directly.
• Survey-flag boxes drawn on your photo, with a tip for every container type.
• Confidence slider that filters results instantly, with no re-scan.
• Copy to-do list for the spots on screen.
• Walk tracker: five photos make a proper sweep.
• Go-outside check: an optional, rule-based nudge from live weather (Open-Meteo), including whether recent rain may have refilled containers. It runs only when you tap it.
• Light mode and a dark "night forest" mode, keyboard accessible, respects reduced motion.
• Friendly loading, empty and error states. No build step, no framework.

How it works

flowchart LR
  A[Phone camera] --> B["index.html<br/>resize to 1280 px"]
  B -->|POST /api/detect| C[FastAPI]
  C --> D["YOLOv8n<br/>best.pt"]
  D -->|labels, boxes, scores| B
  B --> E["Flags, tips,<br/>to-do list"]

Loading

The page shrinks the photo, sends it to POST /api/detect, and draws the returned boxes on a canvas. The server reads the photo in memory, runs the model, and returns JSON. It does not write photos to disk.

Model and results

YOLOv8n, trained for 30 epochs at 640 px on the Breeding Place Detection part of the VisText-Mosquito dataset (3,862 training images, 371 validation, 183 test, five classes). Evaluated on the dataset's held-out test split: 183 images, 424 objects.

Class	Test instances	AP50
Bottle	50	0.884
Coconut-Exocarp	101	0.973
Drain-Inlet	54	0.956
Tire	115	0.893
Vase	104	0.923
All	424	0.926 (mAP50)

Precision 0.906, recall 0.844, mAP50-95 0.692.

Read these numbers carefully. The test photos come from the same collection as the training photos (daylight scenes from Bangladesh). They show the model works on that kind of image. They do not guarantee the same accuracy on every home, street or camera. Confidence scores are not calibrated probabilities, and crowded scenes with many similar objects can score lower.

I also trained a longer run (80 epochs, stopped early at epoch 78). It did not beat the 30-epoch model on the test split (mAP50 0.914 vs 0.926, recall 0.832 vs 0.844), so I kept the original.

Field test

I scanned 5 of my own photos (daylight, around my home) at the default 25% setting. This is a small sample, a sanity check and not a benchmark.

• Tires: 1 of 3 found (2 clear, 1 partly hidden), at 41% confidence.
• False alarms: in the same photo, a parked car was labelled "vase" at 87%, and a vague shape in a dark corner was labelled "coconut shell" at 70%. Both scored higher than the real tire, so raising the confidence slider does not fix this.
• Out-of-scope containers: a dustbin was labelled "vase" at 83%, and paint buckets were labelled "vase" (1 at 25%, 3 at 10%). The label is wrong, but these are real water-holding containers, so the tip still applies.
• Misses: in one photo, 2 terracotta water pots, a plastic bucket and a rooftop water tank got no flags even at 10%. An open drain channel also got nothing.

The model works on the container types it was trained on. It does not cover common household containers like buckets, pots and tanks, and on photos unlike its training data its confidence scores are not a reliable filter.

Experiment: an open-vocabulary model

I also tried YOLO-World (zero-shot, no training) with a word list I typed in. I chose the words after seeing my photos, so this is a demo, not a fair test.

	My fine-tuned YOLOv8n	YOLO-World (zero-shot)
mAP50 on the dataset test split	0.926	0.297
Tire recall	0.80	0.017
Drain-Inlet AP50	0.956	0.000
Parameters when loaded	about 3 million	about 164 million (includes its text encoder)
Saved model file	about 6 MB	329 MB

On my own photos it found a clay pot (89%), a bucket (77%) and a water tank, which my model missed entirely. But it also called a parked car a "water tank" (62%) and a bathroom basin a "water tank" (72%). The two models are good at opposite things, so I kept the small one, which is far more accurate on its five classes. I also checked whether YOLO-World could run in the app. Even with the word list baked in, the saved model is a 329 MB file and used about 1.7 GB of RAM in my Colab test, far over the free server's 512 MB, so I did not deploy it. The next step would be to use YOLO-World to pre-label household containers, fix the labels by hand, and fine-tune the small model on them.

API

POST /api/detect with multipart/form-data:

Field	Type	Notes
file	image	JPG or PNG, up to 15 MB. Resized to fit 1280 px.
conf	float	Minimum confidence, default 0.10, clamped to 0.05 to 0.95.

Example response shape (values are illustrative):

{
  "width": 960,
  "height": 1280,
  "detections": [
    { "label": "Tire", "conf": 0.91, "box": [0.12, 0.40, 0.55, 0.88] }
  ]
}

box is [x1, y1, x2, y2], normalised to 0 to 1 so it can be drawn at any size. GET /health returns {"ok": true}.

Why open source mattered

• A small model for one job. YOLOv8n is trained on exactly five container types and runs on a free CPU instance. There is no per-photo API fee.
• You control where photos go. Detection happens in a server you run, not in a third-party AI service. Self-hosting is one command (below). Detection runs on a server you control. Only the optional weather card and web fonts call out.
• Swappable. Replace best.pt with a retrained model and the app keeps working, because the contract is just labels, scores and boxes.
• Inspectable. Code, weights, dataset and licences are all public, so every claim in this README can be checked.
• The trade-off. Ultralytics YOLO is AGPL-3.0, so this project is released under AGPL-3.0 too.

Run it locally

git clone https://github.com/KK-Dev0/mosquito-breeding-detector.git
cd mosquito-breeding-detector
pip install -r requirements.txt
uvicorn app:app --port 8000

Open http://localhost:8000. Make sure best.pt sits next to app.py.

Deploy on Render

Setting	Value
Build command	pip install -r requirements.txt
Start command	uvicorn app:app --host 0.0.0.0 --port $PORT
Environment variable	PYTHON_VERSION = 3.11.9

The free plan has 512 MB of memory, which is tight for PyTorch plus a model. If the service restarts during a scan, check the logs for an out-of-memory message.

Project structure

app.py             FastAPI server: POST /api/detect, GET /health, serves the page
index.html         The whole frontend (HTML, CSS, vanilla JS), no build step
best.pt            Trained YOLOv8n weights
requirements.txt   CPU-only PyTorch, Ultralytics, FastAPI, Pillow
LICENSE            AGPL-3.0

Limits and responsible use

• It finds containers that can hold water. It cannot tell whether there is water inside, so look in each one.
• It knows five container types. Buckets, water drums, pot saucers and many other breeding sites are not in the training data.
• It is not a medical tool. It does not diagnose, treat or prevent dengue. Follow your local health department's guidance.
• Tips are general advice about emptying, scrubbing, covering and binning containers.
• Confidence scores are not a reliable filter on photos unlike the training data. In my field test, a parked car scored 87% as a "vase" while a real tire scored 41%.

Ideas for next steps

• Classify water present or absent using the dataset's water-segmentation subset.
• Add classes for buckets and water drums, with new labelled photos.
• Use an open-vocabulary model such as YOLO-World to pre-label buckets, pots and water tanks, fix the labels by hand, and fine-tune the small model on them.
• Test and report accuracy on photos from Indian homes.

Credits and licence

• Dataset: Faiyaz, Md Faiyaz Abdullah Sayeedi; Jisan, Md Adnanul (2025). VisText-Mosquito: A Multimodal Dataset for Mosquito Breeding Site Detection, Surface Segmentation, and Reasoning. Mendeley Data, V3. DOI 10.17632/rtsfh7jh7p.3. Licensed CC BY 4.0.
• Model code: Ultralytics YOLOv8, AGPL-3.0.
• Weather: Open-Meteo, used only by the optional go-outside check.
• This project: released under AGPL-3.0. See LICENSE.

Built by Kaushikee Karmakar, BCA, Asansol Engineering College. GitHub: @KK-Dev0.