ShivRamSaud/mushroom-net

Mushroom Net: Large-Scale Normalized CLIP-Style Dataset

Overview

Mushroom Net is a comprehensive, human-curated dataset of mushroom images extracted from 53 unique source archives (~69 GB). It is designed to train high-performance vision models (CLIP, ResNet, ViT) for mushroom classification, edibility prediction, and object detection.

• Total Images: 227,230
• Resolution: 224x224 RGB (Normalized)
• Unique Classes: 2,350+ Normalized Taxonomies
• Classification Heads: 4 (Species, Genus, Subspecies, Edibility)
• Splits: 80% Train / 20% Validation
• Metadata: Comprehensive metadata.csv mapping every image to species, genus, edibility, and source provenance.

Dataset Structure

1. Classification Data (data/)

• train/[class_name]/[class_name]_[idx].jpg: Training images sorted by species.
• val/[class_name]/[class_name]_[idx].jpg: Validation images sorted by species.

2. Detection Data (data_detection/)

• images/: Resized images for object detection.
• annotations_d1.json, etc.: Unified scaled COCO format labels.

2. Metadata (metadata.csv)

The root metadata file contains the catalog of all images:

• final_path: Path from root to image.
• species: Standardized mushroom name.
• genus: Taxon genus (if identified).
• subspecies: Biological variety marker (e.g., var., f. if present).
• edibility: edibility status (edible, poisonous, unknown).
• split: split assignment (train, val).

Quick Start (PyTorch)

Use the provided dataloaders package to load any part of the project:

1. ResNet / CNN Classification

from dataloaders import get_resnet_loader
loader = get_resnet_loader(root_dir="./data", csv_file="./data/metadata.csv", split="train")

2. CLIP (Image-Text Pairs)

from dataloaders import get_clip_loader
loader = get_clip_loader(root_dir="./data", csv_file="./data/metadata.csv", split="train")

3. Multi-Task / Multi-Label Taxonomy

from dataloaders import get_multitask_loader
loader = get_multitask_loader(root_dir="./data", csv_file="./data/metadata.csv", split="train")

4. Object Detection (COCO Format)

from dataloaders import get_detection_loader
loader = get_detection_loader(data_dir="./data_detection", anno_file="annotations_d1.json")

Citation

If you use this dataset in your research, please cite it as:

@dataset{saud2026mushroomnet,
  author = {Shiv Ram Saud},
  title = {Mushroom Net: A Large-Scale Normalized CLIP-Style Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/ShivRamSaud/mushroom-net}
}

Licensing & Attribution

• License: Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC 4.0).
• Usage: This dataset is intended for research and educational purposes. Commercial use is prohibited without express permission.
• Data Sources: Aggregated from 53 unique mushroom archives across multiple open-source biological repositories.

Performance & Training

Mushroom Net is optimized for:

• CLIP: Training robust vision-language encoders.
• Taxonomy: Fine-grained hierarchical classification.
• Object Detection: Precise localization of fungi in natural environments.

Data Cleanup & Reliability

Every image has been sanitized:

• Corrupt Pruning: All zero-byte or unreadable files removed.
• Normalization: Unified 224x224 resizing using Lanczos resampling.
• Deduplication: Repeated archive listings pruned for distinct samples.
• Taxonomic Cleanup: 440+ numeric/internal ID folders unified into 'unidentified'.

Citation & Sources

This dataset aggregates several public mushroom collections. Please refer to individual archive metadata (in data/detection/ and src/pipeline/config.py) for specific archive origins.

---

Created and Standardized by Antigravity AI Assistant