Flame volume is a fundamental descriptor of fire dynamics and is closely related to heat release rate and fire spread potential. Traditional methods for assessing fire size are indirect, relying on measurements of mass loss rate, oxygen consumption calorimetry or simplified geometrical assumptions. With recent advances in computer vision and artificial intelligence (AI), new opportunities arise for quantifying flame geometry directly from visual data. In this study, the semantic segmentation models U-Net, ENet, Fast-SCNN, and two custom DeepLab variants, were trained on datasets from small-scale calibration tests as well as full-scale façade fire experiments. The models were evaluated in terms of segmentation accuracy, inference speed, and generalization ability. Results show that DeepLab-B0 achieved the highest segmentation accuracy, while Fast-SCNN provided the best trade-off between speed and precision. The findings suggest that AI-based segmentation combined with tailored geometric modeling can provide reliable estimates of flame volume, with clear trade-offs depending on whether accuracy or speed is prioritized.