Myoelectric control of hand prostheses requires users to learn unnatural command sequences, which can be both physically and mentally demanding. Semi-autonomous control seeks to automate some of them by anticipating the user's intentions, but it often relies on large, data-heavy computer vision systems that limit real-time use. Our approach applies basic geometric processing to depth data gathered by Time-of-Flight (TOF) sensors to extract object dimensions and orientation. We compared VL53L5CX and VL53L7CX sensors in different setups at various distances from the target. The narrow setup with VL53L5CX (45 mm between sensors) estimated the primary dimension (width/height) and orientation with 15 mm average error and a 66.6 % success rate. VL53L7CX performed better in orientation recognition but showed larger dimension errors and variance. Compact TOF sensor layouts, especially with different fields of view, proved effective for sparse point cloud-based recognition of object properties.