Select Page

Object Detection API

Rationale

To determine the suitability of automated object recognition for conceptual image analysis, this study undertook a comparative evaluation of two widely used models: the Google Vision API and Clarifai’s General Image Recognition model. The context for this evaluation is a broader interdisciplinary research project exploring human developmental needs, visual culture, and symbolic meaning embedded in images. The analysis therefore aimed to assess not only raw detection accuracy, but also semantic coverage, confidence calibration, and the statistical relationship between prediction certainty and correctness.

Methodology

A total of 50 images were selected across the project’s core themes. Each image was manually annotated to establish a ground truth set of expected concepts, including both literal and symbolic indicators relevant to the project’s theoretical framework (e.g., “tree,” “gesture,” “freedom,” “sunset”). Images were then submitted to both APIs, with all returned labels including a model-generated confidence score ∈ [0, 1.0]. Human calibration grounded the evaluation in interpretive accuracy, not just technical recall.

Summary of Analysis

Clarifai was selected as the primary object detection model for this project because it consistently captured a broader and more nuanced range of objects including symbolic and contextual features compared to Google Vision, whose outputs were narrower and more conservative. However, Clarifai’s greater semantic coverage came with an increased risk of high-confidence false positives.

To address this, I conducted statistical threshold analysis using the manually annotated calibration set. Multiple metrics (including precision, recall, F1 score, and Youden’s J) were calculated across a range of confidence thresholds, with supporting significance tests (Mann-Whitney U, paired t-test). This process identified 0.87 as the optimal confidence threshold: at this level, Clarifai achieved a strong balance between precision (91%) and recall (95.5%), while minimising false positives.

By calibrating Clarifai’s output in this way, I ensured that subsequent image analysis was both semantically rich and methodologically robust, aligned with the project’s conceptual aims and empirical requirements.

←→