How koushik9010/nativeanimalbd works
1. Repository purpose
NativeAnimalBD is not an application codebase; it is a curated image dataset repository for Bangladeshi native animals. Its architecture is centered on data organization, distribution, and citation, rather than runtime services or executable components.
2. Core asset model
The primary asset is a hierarchical image corpus of 8,000 images arranged for computer vision and classification experiments. The hierarchy is two-level: 5 taxonomic super-classes and 5 subclasses per super-class, with 400 images per subclass.
3. Top-level dataset partitions
The dataset is split into five main branches:
- Aves
- Mammals
- Insecta
- Reptiles
- Fish
Each branch is meant to act as an independently distributable unit, and the README notes that the complete dataset is published as separate compressed files per main class.
4. Subclass structure and semantic boundary
Within each main class, images are further grouped into species-level subclasses. Examples include:
- Aves: Heron, Kingfisher, Magpie, Sparrow, Parrot
- Mammals: Asian_Elephant, Indian_Leopard, Royal_Bengal_Tiger, Spotted_Deer, Wild_Boar
- Insecta: Ants, Bee, Butterfly, Dragonfly, Grasshopper
- Reptiles: Crocodile, Gharial, Lizard, Snake, Turtle
- Fish: Byen, Koi, Sing, Taki, Tengra
This hierarchy is the key architectural feature for downstream labeling and benchmarking.
5. Data source boundary
The repository itself does not appear to store or generate the images directly in this checkout. Instead, it acts as a distribution and documentation layer for the dataset, while the actual image files are hosted externally on Zenodo. That makes Zenodo the main storage/infrastructure boundary for retrieval.
6. Fish-class exception
A notable architectural constraint is that the Fish images are not redistributed in this repository. Users are instructed to obtain them from the BDINDIGENOUSFISH2019 dataset. This creates a cross-dataset dependency: NativeAnimalBD is partially self-contained, but the full hierarchy requires an external source for one branch.
7. Intended workflow / data flow
The implied data flow is:
Zenodo download → extract class archives → use hierarchical folder labels for training/evaluation.
The folder structure is the authoritative label source, so consumers likely rely on directory names rather than a separate annotation file.
8. Research and benchmarking role
The dataset is designed for hierarchical image classification, computer vision, deep learning, and biodiversity informatics. Architecturally, this means the hierarchy is intentional: it supports experiments where models predict either coarse classes, fine subclasses, or both levels.
9. Distribution and reproducibility
The dataset is versioned and citable through Zenodo DOI 10.5281/zenodo.18729571, which is the stable distribution endpoint. This is important for reproducibility: experiments can reference a persistent artifact rather than an ad hoc file share.
10. Licensing boundary
The repository documents that images were collected from Unsplash and Pexels for non-fish classes, and users must comply with those platforms’ licenses. So the legal boundary is split by source: the dataset is a curated aggregation, not a single-origin asset pack.
11. Repository content scope
Based on the tree provided, the repo currently contains only the README documentation. That means the repository’s architectural surface is minimal: it defines the dataset schema, distribution instructions, provenance, and citation metadata, but not preprocessing code, training code, or annotation tooling.
12. What to draw in an architecture diagram
An accurate diagram should show:
- a documentation/metadata layer in the repo
- an external Zenodo distribution store
- the hierarchical dataset structure as the central data model
- an external dependency for Fish via BDINDIGENOUSFISH2019
- consumers such as ML training/evaluation workflows
The main boundary is between the lightweight repository documentation and the externally hosted image archives it describes.
Main entry points: README.md README
Components 16
Repository
Lightweight documentation and metadata layer.
- README · Primary repository content and dataset instructions.
README.md
Dataset model
Hierarchical image corpus organized for classification.
- NativeAnimalBD · Curated 8k-image corpus with hierarchical labels.
- 5 classes · Top-level taxonomy used as the main partitioning scheme.
- Aves · Bird branch; distributed as its own archive.
- Mammals · Mammal branch; distributed as its own archive.
- Insecta · Insect branch; distributed as its own archive.
- Reptiles · Reptile branch; distributed as its own archive.
- Fish · Fish branch; not redistributed in this checkout.
- Species folders · Second-level species labels under each main class.
External sources
Hosting, provenance, and cross-dataset dependencies.
- Zenodo · Stable distribution endpoint for the dataset archives.
- DOI · Persistent citation for the published dataset artifact.
- BDINDIGENOUSFISH2019 · Required source for the fish branch.
- Unsplash · Image provenance source for non-fish classes.
- Pexels · Image provenance source for non-fish classes.
Consumers
Downstream research and benchmarking workflows.
- Training · Uses folder names as authoritative labels for model training.
- Benchmarking · Supports hierarchical classification experiments and evaluation.
Connections 22
- README → NativeAnimalBD: describes · The README defines the dataset structure and usage.
- README → Zenodo: download · Readers are directed to the Zenodo archive for retrieval.
- README → DOI: cites · The repository documents a persistent DOI for the release.
- README → Unsplash: provenance · Non-fish images are collected from Unsplash.
- README → Pexels: provenance · Non-fish images are collected from Pexels.
- NativeAnimalBD → 5 classes: organized as · The corpus is arranged into five super-classes.
- 5 classes → Aves: includes · Bird branch.
- 5 classes → Mammals: includes · Mammal branch.
- 5 classes → Insecta: includes · Insect branch.
- 5 classes → Reptiles: includes · Reptile branch.
- 5 classes → Fish: includes · Fish branch.
- Aves → Species folders: refines · Species-level folders sit beneath the branch.
- Mammals → Species folders: refines · Species-level folders sit beneath the branch.
- Insecta → Species folders: refines · Species-level folders sit beneath the branch.
- Reptiles → Species folders: refines · Species-level folders sit beneath the branch.
- Fish → BDINDIGENOUSFISH2019: depends on · Fish images are obtained from the external fish dataset.
- Zenodo → NativeAnimalBD: hosts · Primary distributed archives are published on Zenodo.
- DOI → Zenodo: resolves to · The DOI identifies the archived dataset release.
- Training → NativeAnimalBD: consumes · Models use the folder hierarchy as label input.
- Benchmarking → NativeAnimalBD: evaluates · The hierarchy supports coarse- and fine-grained experiments.
- Benchmarking → 5 classes: uses · Coarse classification can target the top-level classes.
- Benchmarking → Species folders: uses · Fine-grained classification can target species folders.