PlatformMarketplaceSolutionsResourcesOpen DatasetsCommunityCompany
update dataset overview and ba...
Feb 10, 2022 7:31 AM


AnimalWeb has been constructed following the animal biological taxonomy. It populates faces from 334 different species spread over 21 different animal orders. The proposed dataset is aimed at offering a large-scale and diverse coverage of annotated animal faces. It contains 21.9K annotated faces, offering 334 different animal species with variable number of animal faces in each species.
29% of the total species contain 65% of the total faces. Also, the maximum and minimum number of faces per specie are 241 and 1, respectively. Both these statistics highlight the large imbalance between species and high variability in the instance count for different species. This marks the conformity with the real-world where different species are observed with varying frequencies.

Data Collection

Image Collection

• Preparation of a diverse and extensive taxonomic data structure
• Preparation of a detailed data collection protocol to ensure realworld conditions
• A team of 3 trained volunteers under the supervision of an expert completed the collection process. For each worker, it took an average of 100 images/hour. [~250 man-hours]
• Visual filtering step to avoid potential duplicates across every species. [43.8 man-hours]

Workflow Development

• Project review and approval by a panel of "zooniverse" citizen science experts
• Metadata prepared and loaded to server
• Workflow is designed for annotating 9 pts to be easily usable for volunteers of various domain expertise. “Order” and “name” for each facial point defined.
• Clear action-plan in case of ambiguities (e.g., invisible landmarks)
• Workflow linked with a professionally developed help page showing instructions and illustrations to annotate all possible species across diverse poses.
• Workflow thoroughly tested by a 5- member expert team. [20 man-hours]

Data Annotation

Facial point annotation

• Zooniverse volunteers have a prior experience of annotating many different successful citizen science projects related to animals.
• Every face is annotated by at least 5 different volunteers. [~5408 manhours]
• The annotation portal allows annotators to raise a query with the experts throughout the annotation life cycle.
• The whole exercise of zooniverse crowdsourcing took 80 man-hours of experts’ time.

Refining annotations

• A team of 4 members hired and trained for refinement. • Team supervised by an expert [45 manhours] • In the first stage, major errors were rectified e.g., correcting points ordering This refinement proceeded species wise to enforce consistency in annotations across possible every species. [548 manhours] • In the second stage: pixel perfect annotations were ensured by cross-annotator review. [438 man-hours]


      title={AnimalWeb: A Large-Scale Hierarchical Dataset of Annotated Animal Faces},
      author={Muhammad Haris Khan and John McDonagh and Salman Khan and Muhammad
 Shahabuddin and Aditya Arora and Fahad Shahbaz Khan and Ling Shao and Georgios Tzimiropoulos},
🎉Many thanks to Graviti Open Datasets for contributing the dataset
Basic Information
Application ScenariosNot Available
AnnotationsNot Available
TasksNot Available
Updated on2022-02-10 07:31:03
Data TypeNot Available
Data Volume0
Annotation Amount0
File Size0.00B
Copyright Owner
Muhammad Haris Khan