JabutiVisionData - Annotated Image Dataset for Morphological Differentiation of Chelonoidis carbonaria and Chelonoidis denticulata

The JabutiVisionData dataset is an annotated image collection designed to support computer vision research focused on the morphological differentiation between Chelonoidis carbonaria (red-footed tortoise) and Chelonoidis denticulata (yellow-footed tortoise). These two species exhibit high morphological similarity and partial geographic overlap in South America, making visual identification challenging, particularly in fine-grained classification scenarios. The dataset consists of 1,150 RGB images collected from the citizen science platform iNaturalist, representing photographic records captured in natural environments. A rigorous manual curation process was applied to ensure taxonomic correctness, remove duplicate records, and exclude images with insufficient visual quality. The curated dataset contains a total of 1,173 annotated objects, distributed in a balanced manner between the two species. Annotations were performed manually using bounding boxes that fully encompass the visible portions of each individual, including the carapace, limbs, and head when present. Images containing multiple individuals were annotated with independent bounding boxes for each specimen. In cases of partial occlusion, only clearly visible and identifiable regions were annotated to avoid inference of non-visible areas. All annotations were reviewed to ensure spatial consistency and label accuracy.
The dataset is provided in both COCO and YOLO annotation formats, ensuring interoperability with a wide range of deep learning frameworks and object detection architectures. The data are partitioned into training (74%), validation (14%), and test (12%) subsets, preserving class balance across all splits. Statistical analyses of bounding box dimensions reveal substantial variability in object scale and aspect ratio, reflecting realistic variations in capture distance, pose, and framing. JabutiVisionData is suitable for training and evaluating object detection and classification models, benchmarking deep learning architectures under fine-grained visual discrimination conditions, and supporting non-invasive environmental monitoring applications. Its structure also enables integration with camera trap systems and edge AI devices, facilitating real-world biodiversity monitoring and ecological research.

To organize the dataset clearly and professionally for YOLOv12 training, the folder structure should be arranged following the COCO/YOLO standard as follows:

JabutiVisionData/
├── test/
│ ├── _annotations.coco.json # train metadata
│ └── images/ # 141 images
├── train/
│ ├── _annotations.coco.json
│ └── images/ # 849 images
└── valid/
├── _annotations.coco.json
└── images/ # 160 images