The ORBIT India Dataset: Understanding the Challenges of Collecting a Disability-First AI Dataset in Low-Resource Environments
Gesu India, Martin Grayson, Cecily Morrison, Daniela Massiceti, Simon Robinson, Jennifer Pearson, Matt Jones · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3791099
Summary
This paper introduces ORBIT-India, the first teachable object recognition dataset contributed entirely by people who are blind or have low vision in India. It extends the UK/Canada-collected ORBIT dataset (Massiceti et al., 2021) to the Indian context — home of the world's largest blind population — to address a glaring gap: computer vision systems used by blind people are overwhelmingly trained on Global North data that fails to reflect the material culture, lighting, backgrounds, and object varieties of Global South homes. The authors adapted Microsoft's Find My Things app (originally iOS) to Android, localised it into Hindi and Indian English, and worked with 12 blind or low-vision data collectors (nine men, three women, average age 27) across tier-I Indian cities between July and December 2023. Each collector selected around ten personally relevant objects and recorded eight short videos per object (six training, two testing) following the Phase II ORBIT protocol. Beyond the dataset itself, the paper is a reflexive methodological contribution. The authors address three research questions: how a Global-North dataset collection protocol must be adapted for low-resource Global South contexts; what the collected data actually looks like in terms of visual, contextual, environmental, and data-collector experiences; and what lessons emerge for cross-geography AI dataset collection. They ground the work in data feminism (D'Ignazio and Klein), data colonialism critiques (Couldry and Mejias), and the disability-first framing established by the original ORBIT project. Methods combine dataset statistics, comparative analysis with the parent ORBIT dataset, and reflexive thematic analysis of ten post-collection semi-structured interviews.
Key findings
ORBIT-India comprises 105,243 images extracted from 587 videos of 76 objects, with an average of 6.33 objects contributed per data collector (versus eight in the original ORBIT). The dataset surfaces culturally distinctive objects largely absent from Global North datasets: bangles, hair oil, trofi (trophy), belan (rolling pin), steel glasses, and Indian currency notes. Household organisation also differs — toothbrushes stored away from bathroom sinks in joint-family homes, for instance — giving models exposure to spatial patterns they would otherwise miss. Only about 1.11% of frames (1,170 images) had the object completely out of frame, and 3.80% contained PII — markedly lower than VizWiz's ~13% — attributed to careful participant behaviour and protocol guidance. Data collectors spent an average of 3.5 hours per participant (up to 5), with single-object recording taking ~5 minutes including scene setup. Six of 12 reported battery drain and phone heating from continuous camera use; the Android port introduced lower camera quality and limited accessibility APIs compared to iOS. Qualitative findings show participants exercised nuanced judgment about privacy — one collector knowingly shared trophies displaying workplace addresses because the objects symbolised pride, while another avoided filming keys entirely. Perceptions of what counted as 'private' were culturally embedded (e.g., living rooms as semi-public guest space). Data collectors valued in-app filming instructions and haptic/audio feedback but requested a 'lighting clarity percentage bar' and customisable feedback. Recruitment favoured urban, male, digitally literate collectors — rural, female, and older blind Indians were underrepresented.
Relevance
For accessibility practitioners and AI teams, this paper is a practical template for extending disability-first AI tools into underrepresented geographies. It shows concretely what breaks when Global-North dataset protocols are lifted into Global-South contexts: device ecosystems (iOS vs Android), connectivity (GB-scale uploads), language (multilingual instructions and labels), recruitment channels (online lists miss rural participants), and even legal regimes (India's DPDPA 2023 lacks GDPR's Right to be Forgotten). The authors argue persuasively that small, high-quality, culturally situated datasets can meaningfully improve few-shot recognition even without matching VizWiz-scale volume — a useful counterweight to 'bigger is always better' dataset thinking. Limitations are openly named: 12 urban collectors cannot represent India's rural, linguistic, and gender diversity, and tools like ARCode constrain participation to mid-to-high-range Android phones. Practitioners building object recognition, image description, or visual assistance features should treat this as required reading before claiming global coverage.
Tags: AI · accessibility · datasets · teachable object recognition · vision impairment · Global South · blindness and low vision · computer vision · few-shot learning · data feminism · data colonialism · participatory design · privacy · PII · India
Standards referenced: GDPR · DPDPA 2023