README.md

---
license: cc-by-4.0
task_categories:
- object-detection
- image-segmentation
language:
- en
pretty_name: ShitSpotter
---
# Dataset Card for ShitSpotter ("ScatSpotter")


ShitSpotter (or "ScatSpotter" in formal settings) is an open dataset of images containing dog feces. 
This dataset contains full-resolution smartphone images of dog feces ("poop") collected in urban outdoor 
environments taken using a "before/after/negative" protocol. 
It includes thousands of polygon annotations of feces in varied lighting, seasonal, and terrain conditions. 
The dataset is designed for training and evaluating object detection and segmentation models, 
with potential applications in public sanitation, AR-assisted navigation, and pet owner assistance.

This dataset card was generated via ChatGPT (based on existing documentation) and edited by the author.

## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->

Images were collected using a smartphone and annotated with manually or AI-assisted drawn polygons using the Segment Anything Model (SAM). 
The goal is to support the development of systems that can detect and segment dog feces in real-world environments.

The dataset is publicly released with open-source licenses and is distributed via IPFS, BitTorrent,
a centralized Girder server, and HuggingFace.


- **Curated by:** Jonathan Crall
- **Funded by:** Not funded; this is a self-motivated project.
- **Shared by:** Kitware (institutional affiliation of the author)
- **License:**, Code: Apache 2.0, Data: Creative Commons Attribution 4.0 International (CC BY 4.0)


### Dataset Sources 

- **Repository:** https://github.com/Erotemic/shitspotter
- **Paper:** Preprint: https://www.arxiv.org/abs/2412.16473
- **Demo:** No live demo at this time. (Attempting a react-native-vision-camera app, but progress is slow. Help wanted.)


## Uses


### Direct Use

The dataset is intended for training, evaluating, and benchmarking computer vision algorithms, particularly:

* Object detection and segmentation (single-class)
* Dataset analysis, robustness testing

Trained models could enable:

* Deployment in mobile or AR applications for locating feces
* Urban cleanliness monitoring by municipalities
* AR poop collision warning systems

### Out-of-Scope Use

The images are high resolution, but the poops themselves sometimes are not. 
There are images from multiple scales, but the majority occupy smaller areas of the image.
Thus more fine-grained analysis of the poops themselves may not be possible.

The dataset was not collected or labeled with health diagnostics in mind, 
so caution should be used when applying it to medical contexts. 
However, it does contain healthy and sick poops, the data may still offer
useful signals for future health-related research if properly validated.

As with any open dataset, respectful use is expected. 
This dataset was created to advance research and practical applications 
in public health, computer vision, and sanitation.


## Dataset Structure

The dataset consists of RGB JPEG images (mostly 4032×3024), many with EXIF metadata and kwcoco-format metadata. 
Most data follows a “before/after/negative” triplet protocol:

“Before” image: includes visible feces

“After” image: same scene after cleanup (used as high-correlation negative)

“Negative” image: a different, nearby image without feces (used as low-correlation negative)

Only the “before” images contain annotations. Annotations are polygons in COCO-style JSON format. 

Data is split into:

* Train: Author collected images from 2021–2023 and 2/3 of 2024, and 2025+.

* Val: Author collected from 2020 and 1/3 of 2024

* Test: 121 contributed images with annotations, held out for evaluation


## Dataset Creation

### Curation Rationale

The dataset began after a real-world frustration: the creator couldn’t find their dog’s poop in a leaf-covered park. 
This sparked the idea of building a vision model to solve the problem and contribute a novel, underexplored dataset to the ML community.
The dataset serves practical, research, and humorous goals. 
It also explores data distribution over IPFS and the lifecycle of living datasets.


### Source Data

#### Data Collection and Processing

Collected with a Pixel 5 and other smartphones

Annotation tool: LabelMe

AI-assistance: Segment Anything Model (SAM), trained detectron models (always manually reviwed / edited).

Preprocessing includes optional alignment of image pairs (not always reliable)

#### Who are the source data producers?

The vast majority of images were taken by the dataset curator. 
A few contributors have added images, which are separately labeled. 
Images were taken during normal dog walks in parks and sidewalks.

### Annotations

#### Annotation process

Annotations were created in LabelMe. 
Some were SAM-initialized using click prompts and corrected manually. 
Some were seeded with a detectron model. 
Some were manually drawn.
Shadows and similar-looking natural objects are known challenges.
All annotations are manually reviewed.

#### Who are the annotators?

The dataset creator; contributors may annotate in future.

#### Personal and Sensitive Information

Some EXIF location metadata is stripped, but many images are unmodified.
Consent was received to publish EXIF metadata along with these images.

## Bias, Risks, and Limitations

Geographic bias: Mostly Upstate NY.

Sensor bias: Primarily Google Pixel 5

Collector/dog bias: Same person and 3 dogs over time

Bias in freshness, specific dogs, and lighting: More recent, fresh, daytime
images with poops from the author's dogs dominate.


### Recommendations

When applying models trained on this dataset to new domains (e.g., different locations, animals, lighting), 
expect a performance gap unless domain adaptation is applied. 
Future work may include generalizing across feces types.


## Citation

The current paper is not peer-reviewed. When a peer-reviewed version of the
paper is published we will update the citation.

**BibTeX:**

```
@article{crall2024scatspotter,
  title={"ScatSpotter" 2024 — A Distributed Dog Poop Detection Dataset},
  author={Crall, Jonathan},
  journal={arXiv preprint arXiv:2412.16473},
  year={2024}
}
```

**APA:**

Crall, J. (2024). "ScatSpotter" 2024 — A Distributed Dog Poop Detection Dataset. arXiv:2412.16473.


## Glossary

B/A/N Protocol: Before/After/Negative triplets for contrastive learning

IPFS: InterPlanetary File System, a decentralized, content-addressable data store

Polygon annotation: High-precision mask of object shape


## More Information 

Dataset access via: IPFS, BitTorrent, and Girder

Datasheet: Available in repo

2024 Paper Preprint: https://www.arxiv.org/abs/2412.16473

Torrent: https://academictorrents.com/details/ee8d2c87a39ea9bfe48bef7eb4ca12eb68852c49

IPNS address: /ipns/k51qzi5uqu5dje1ees96dtsoslauh124drt5ajrtr85j12ae7fwsfhxb07shit

Recent IPFS CID: /ipfs/bafybeihsd6rwjha4kbeluwdjzizxshrkcsynkwgjx7fipm5pual6eexax4

Girder Mirror: https://data.kitware.com/#user/598a19658d777f7d33e9c18b/folder/66b6bc34f87a980650f41f90

## Dataset Card Contact

GitHub Issues: https://github.com/Erotemic/shitspotter/issues

Email: [email protected]