💻 computer science

"ScatSpotter" -- A Dog Poop Detection Dataset

The paper introduces "ScatSpotter," a new dataset of over 9,000 images featuring polygon-annotated dog feces collected via a unique "before/after/negative" protocol to advance the detection and segmentation of small, camouflaged outdoor waste.

Original authors: Jon Crall

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jon Crall

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The "ScatSpotter" Project: Teaching Robots to Spot the Unspottable

Imagine you are walking through a beautiful, leafy park on a sunny afternoon. You’re looking at the trees, the birds, and the flowers. But there is one thing you are really looking for: dog poop.

It’s a gross thought, but it’s a real problem. If you’re a dog owner, you don't want to step in it. If you’re a city worker, you need to clean it up. And if we ever want "smart" robot vacuums to clean our sidewalks or autonomous drones to monitor our parks, those machines need to be able to see that tiny, brown, camouflaged object hidden under a pile of autumn leaves.

The problem? For a computer, finding dog poop is like trying to find a single specific grain of sand in a desert, or a single brown leaf in a pile of brown leaves. It’s small, it’s oddly shaped, and it blends in perfectly with the "clutter" of nature.

That is why a researcher named Jonathan Crall created "ScatSpotter."

1. The "Before, After, and Maybe" Method (The BAN Protocol)

To teach a computer, you can't just show it a picture and say, "This is poop." You have to be much more clever. Crall used a special method called BAN:

Before: A photo of the poop sitting in the grass.
After: A photo of the exact same spot after the poop has been cleaned up.
Negative: A photo of a nearby area that looks similar (maybe a pinecone or a weirdly shaped rock) but isn't poop.

The Analogy: Think of it like teaching a child to recognize a specific toy. You show them the toy (Before), you show them the empty spot where the toy used to be (After), and then you show them a different toy that looks similar so they don't get confused (Negative). This helps the computer learn the difference between "actual waste" and "just a random stick."

2. The "Brain" Test (The Models)

The researchers took this massive collection of photos (over 9,000 of them!) and fed them into different types of "AI brains" to see which one was the best at the game.

The "Zero-Shot" Brain: This is like a student walking into an exam having never studied the subject. They try to guess based on general knowledge. It performed poorly—it just couldn't "get" what poop looks like.
The "Tuned" Brain: This is the straight-A student who spent all night studying the ScatSpotter textbook. This was the winner! It was much better at drawing a digital "fence" (a polygon) around the poop to show exactly where it was.

3. The "Digital Delivery" Experiment

The paper also asks a very practical question: How do we share this massive amount of data with other scientists?

If you want to send a huge heavy box to a friend, you have two choices:

The Centralized Way (The UPS/FedEx approach): You send it to one big warehouse (like HuggingFace). It’s super fast, but if the warehouse closes or gets too expensive, you're in trouble.
The Decentralized Way (The "Neighborhood Potluck" approach): You break the box into tiny pieces and have hundreds of neighbors hold onto a piece (like BitTorrent or IPFS). It’s much harder to "lose" the data this way, but it can be a lot slower to get all the pieces back together.

The researchers found that while the "FedEx" way is much faster, the "Potluck" way is a great backup to make sure scientific data stays available forever.

Why does this matter?

While "ScatSpotter" is specifically about dog poop, it’s actually a training ground for much bigger things. If a robot can learn to spot a tiny, camouflaged piece of poop, it can eventually learn to spot:

Microtrash (like tiny bits of plastic in the ocean).
Wildlife droppings (to help scientists track endangered animals).
Litter in our streets to keep our cities clean.

In short: ScatSpotter is teaching the machines of the future how to see the small, messy details that humans—and robots—often miss.

Technical Summary: “ScatSpotter” — A Dog Poop Detection Dataset

Authors: Jonathan Crall (Kitware)
Venue: Accepted to the International Workshop on Smart Waste Monitoring (WasteVision) at WACV 2026.

1. Problem Statement

Modern computer vision has made significant strides in detecting large, conspicuous objects. However, detecting small, amorphous, and camouflaged waste—such as biological droppings (dog feces) or microtrash—remains a major challenge. These objects are often low-contrast, irregular in shape, and visually similar to natural environmental clutter (leaves, sticks, rocks), making them difficult for both humans and autonomous systems to identify. This gap hinders applications in autonomous cleanup, public health monitoring, and wildlife surveillance.

2. Methodology

The authors introduce ScatSpotter, a specialized dataset designed to address the difficulty of detecting small-scale outdoor waste.

Data Collection (BAN Protocol): To provide high-quality training and counterfactual examples, the authors employed a “Before/After/Negative” (BAN) protocol. For a given location, they capture:
1. Before: An image with the object present.
2. After: An image of the same scene/viewpoint after the object is removed.
3. Negative: A nearby scene containing visually similar "confusers" (e.g., pine cones or leaves).
Dataset Composition: The dataset contains over 9,000 full-resolution images (4,032 × 3,024) and 6,000 polygon annotations. It features significant variation in lighting, season (snow, sun, rain), and background clutter.
Data Splits:
- Training/Validation: Author-collected images, split by year and capture date to prevent temporal bias.
- Test Set: A 121-image set provided by community contributors, ensuring the model is tested on images from different photographers, devices, and locations to measure true generalization.
Baseline Modeling: The authors evaluated seven model variants, including:
- Segmentation: VIT-sseg-s (trained from scratch).
- Detection: MaskRCNN (pretrained and from scratch), YOLO-v9 (pretrained and from scratch), and Grounding DINO (zero-shot and fine-tuned).

3. Key Contributions

A Novel Dataset: The largest publicly available collection of annotated dog poop images, specifically optimized for small-object detection and segmentation in cluttered environments.
Baseline Benchmarks: A comprehensive suite of trained models that establish performance standards for this specific domain.
Distribution Study: A comparative analysis of centralized (HuggingFace, Girder) vs. decentralized (BitTorrent, IPFS) data distribution mechanisms, evaluating the trade-offs between speed, cost, and scientific reproducibility (content-addressability).

4. Results

Model Performance:
- Zero-shot models (Grounding DINO) performed poorly, highlighting that foundational models currently lack sufficient coverage for this specific category.
- Fine-tuned models performed best; specifically, tuned Grounding DINO achieved a box-level Average Precision (AP) of 0.69 on validation and 0.70 on the test set.
- Pretrained models (MaskRCNN-p, YOLO-v9-p) significantly outperformed those trained from scratch, confirming the necessity of transfer learning for this domain.
Distribution Efficiency: HuggingFace was the fastest distribution method due to optimized CDNs, while decentralized methods (IPFS/BitTorrent) were slower due to peer-discovery overhead but offered superior data integrity and resistance to "link rot."
Resource Estimation: The authors quantified the environmental impact, noting that training the VIT models was the most resource-intensive phase, consuming significant energy and emitting substantial $CO_2$ .

5. Significance

ScatSpotter serves as a prototypical benchmark for the broader problem of detecting small, camouflaged waste. While the dataset focuses on dog feces, the technical challenges it presents—low contrast, irregular shapes, and environmental camouflage—are directly applicable to:

Autonomous Waste Management: Robotic systems for cleaning sidewalks and parks.
Environmental Monitoring: Tracking wildlife via droppings or identifying microtrash in ecosystems.
Augmented Reality: Warning systems in smart glasses to prevent pedestrians from stepping in waste.

By providing high-resolution, polygon-annotated data and a robust testing protocol, the authors provide the community with the tools necessary to advance the state of the art in small-object environmental computer vision.