README.md

metadata

license: mit
task_categories:
  - object-detection
tags:
  - pest-detection
  - coco
  - synthetic
  - surveillance
  - bounding-box
  - mouse-detection
  - rat-detection
  - cockroach-detection
  - kitchen
  - indoor
  - video-frames
  - synthetic-data
  - places365
  - food-safety
  - pest-control
  - computer-vision
  - temporal
  - detr
  - yolo
size_categories:
  - 10K<n<100K

Pest Detection Dataset (Synthetic, COCO Format)

A synthetic dataset for pest detection in kitchen/indoor surveillance scenarios. All videos and annotations are procedurally generated using a custom pipeline.

Demo — Job 64d68645 (1 mouse · 1 rat · 1 cockroach)

Bounding boxes: 🔴 mouse   🟢 rat   🔵 cockroach

---

Dataset Statistics

Split	Jobs	Frames	Annotations	🐭 Mouse	🐀 Rat	🪳 Cockroach	Empty Jobs
train	1,460	36,500	55,650	16,200	10,975	28,475	368 (25.2%)
val	298	7,450	11,650	3,225	2,375	6,050	76 (25.5%)
test	308	7,700	11,575	3,575	2,025	5,975	84 (27.3%)
total	2,066	51,650	78,875	23,000	15,375	40,500	528 (25.6%)

---

Pest Categories

ID	Name	Description
1	mouse	House mouse
2	rat	Common rat
3	cockroach	Cockroach

---

Dataset Structure

├── annotations/
│   ├── train.json       # COCO-format annotations for train split
│   ├── val.json         # COCO-format annotations for val split
│   └── test.json        # COCO-format annotations for test split
├── images/
│   ├── train/
│   │   └── {job_id}/    # One folder per video job (~25 frames each)
│   │       ├── frame_0001.png
│   │       ├── frame_0010.png
│   │       └── ...
│   ├── val/
│   │   └── {job_id}/
│   └── test/
│       └── {job_id}/
├── generated_state.json  # Full metadata for all 2,066 generated jobs
└── demo.mp4              # Demo video with bounding box overlays

Why job-ID subfolders?

Frames from the same synthetic video are grouped under one {job_id} folder. This preserves temporal structure — consecutive frames in a folder belong to the same video sequence — enabling temporal models and frame-sequence dataloaders without needing extra metadata.

---

Annotation Format (COCO)

Each annotation JSON follows the standard COCO object detection format:

{
  "images": [
    {
      "id": 1,
      "file_name": "64d68645/frame_0001.png",
      "width": 640,
      "height": 480
    }
  ],
  "annotations": [
    {
      "id": 1,
      "image_id": 1,
      "category_id": 1,
      "bbox": [x, y, width, height],
      "area": 1234.5,
      "iscrowd": 0
    }
  ],
  "categories": [
    {"id": 1, "name": "mouse"},
    {"id": 2, "name": "rat"},
    {"id": 3, "name": "cockroach"}
  ]
}

Note: file_name includes the {job_id}/ prefix — prepend images/{split}/ to get the full path.

---

Loading the Dataset

from huggingface_hub import snapshot_download
import json, os
from PIL import Image

# Download the dataset
dataset_dir = snapshot_download("adR6x/pest_detection_dataset", repo_type="dataset")

# Load train annotations
with open(os.path.join(dataset_dir, "annotations", "train.json")) as f:
    coco = json.load(f)

# Load an image
img_info = coco["images"][0]
img_path = os.path.join(dataset_dir, "images", "train", img_info["file_name"])
img = Image.open(img_path)

---

generated_state.json

Each entry in generated_state.json describes one generated video job:

{
  "job_id": "64d68645",
  "split": "train",
  "mouse_count": 1,
  "rat_count": 1,
  "cockroach_count": 1,
  "length_of_video_seconds": 24.0,
  "fps": 10,
  "frames_dir": "images/train/64d68645",
  "labels_dir": "annotations/train.json",
  ...
}

Use this to filter jobs by pest composition, split, or other metadata.

---

Data Sources

Asset	Source	Details
Kitchen backgrounds	Places365 + manually curated	70 kitchen images selected from Places365 to cover diverse kitchen layouts, lighting conditions, and viewpoints
Pest models & animation	Custom synthetic pipeline	3D pest models (mouse, rat, cockroach) procedurally animated and composited onto background images
Bounding box annotations	Auto-generated	Derived directly from the synthetic compositing process — no manual labelling required

---

Frame Sampling

Frames are sampled at every 10th frame from the original video (--every_n 10 default), plus frame 1. Each job contributes ~25 frames at 640×480 resolution.