Live flock counts and loss alerts for poultry farms

Flock loss is invisible until it is large. Farms count birds by walking the barn - slow, error-prone and done a few times a day at best - so a problem that starts at 3 a.m. is usually found the next morning, after it has already cost money.

• Count hundreds of overlapping birds accurately from camera snapshots
• Alert the right people automatically when a count falls below its threshold
• Run reliably on hardware a farm can actually afford

We chose density regression over object detection. A DFCCNet model with a ResNet50-FPN backbone predicts a density map whose sum is the count, which keeps working when birds overlap and box-based detectors fall apart. It runs as an isolated service behind a simple contract: an image in, a count and a confidence score out.

A scheduler polls every active camera on a fixed interval, handles both Digest and Basic camera authentication, and checks that each response really is an image before it reaches the model. One failing camera never stops the rest of the sweep.

Every count is stored against its area and checked against a threshold. A breach notifies the farm owner and every assigned staff member by SMS and in-app message - deduplicated, with a delivery record for each person.

1. 01

   Density maps, not bounding boxes

   Counting from a density map is what makes a crowded barn countable at all.

2. 02

   Validate before you predict

   A misconfigured camera URL returns a web page, not a photo. Checking every frame first means bad feeds fail loudly instead of producing a confident, wrong number.

3. 03

   Sized on measurements, not guesses

   We measured about 3.2 GB of peak memory and 7.9 seconds per image on 2 vCPUs, capped images at 600px and kept one inference in flight - so the whole system runs on a single commodity server.

4. 04

   Alerts with a paper trail

   Every alert records the count, the threshold and whether each SMS was sent, failed or had no phone number to go to.

Molt turns flock counting from a scheduled chore into constant observation. Every area has a timestamped count history, mortality shows up as a trend instead of a discovery, and staff are told within one capture cycle - while there is still time to act. Because it runs on one server, the monitoring costs less than the losses it prevents.

Automatic counts for every barn area

24/7

From a drop in count to an SMS alert

1 cycle

Runs the entire system

1 server

How does Molt count birds?

It uses a density-regression model (DFCCNet with a ResNet50-FPN backbone). Instead of drawing a box around each bird, it predicts a density map whose sum is the count, which stays accurate when hundreds of birds overlap.

Which cameras does Molt work with?

Standard IP cameras that serve snapshots over HTTP, including Hikvision and Dahua models that use Digest authentication and cheaper cameras that use Basic authentication.

Who gets alerted when a count drops?

The farm owner and every active staff member assigned to that farm, by SMS and in-app notification. Nobody is messaged twice, and each delivery is recorded.

What does Molt run on?

A single commodity server. The model runs on CPU, images are resized to 600px so memory stays predictable, and one inference runs at a time.

What is Molt built with?

Python, FastAPI and PyTorch for the backend and model, PostgreSQL for data, React and TypeScript for the web app, and Docker for deployment.