Loka Wicara: Technical Evaluation of Local Object Detection Based on YOLOv8 Nano for an AAC Prototype

Abstract

Children with Autism Spectrum Disorder (ASD) may experience functional communication barriers, while Augmentative and Alternative Communication (AAC) applications should use objects and messages that are closely related to users’ daily contexts. This study presents an initial technical evaluation of Loka Wicara, an AAC prototype that links local-object detection to Indonesian need-message designs. The actual dataset contains 500 images and 520 bounding boxes across five classes: pillow, dipper, glass/cup, cooked rice, and sandals. The data were divided into 350 training, 100 validation, and 50 test images. YOLOv8n and a YOLOv5n-compatible model were trained for 100 epochs using the same three random seeds. On the test set, YOLOv8n achieved a precision of 0.8475 ± 0.0652, recall of 0.8267 ± 0.0837, F1-score of 0.8328 ± 0.0241, mAP50 of 0.8768 ± 0.0338, and mAP50–95 of 0.6115 ± 0.0358. YOLOv5n produced slightly higher mAP50 and mAP50–95 values of 0.8800 ± 0.0240 and 0.6146 ± 0.0469, respectively, while using fewer parameters and a smaller model file. A Tesla T4 benchmark produced mean latencies of 18.21 ms for YOLOv8n and 18.69 ms for YOLOv5n. YOLOv8n was also successfully exported to LiteRT/TFLite. The findings indicate that both models are competitive; YOLOv8n is retained as a prototype candidate because of its slightly higher recall and F1-score and its successful export pipeline, rather than universal superiority. Android testing, expert validation, and user acceptance testing remain necessary in subsequent work.