SoloFinger: Single-Finger Grasping Microgestures
SoloFinger microgestures are conceptually based on the observation that during everyday hand-object interaction, multiple fingers tend to move concurrently, whereas it is rare that a single finger moves extensively on the object while all others stay idle. This observation was informed by findings that finger movements tend to be highly correlated during object manipulation. Our work leverages this phenomenon. We ground our findings on objects that do not contain movable elements, such as mechanical buttons or sliders. A SoloFinger gesture (see image below) involves moving a single finger by a considerable yet comfortable extent, while all other fingers remain static. It is performed while holding an object, with the same hand, and on the object itself. SoloFinger gestures are not limited to any specific finger. We recommend using the thumb, index, or middle finger, as the ring and pinky fingers were shown to be less robust and also subjectively less preferable

Validating the SoloFinger Concept with a White-Box Classifier
To validate this concept, we record the finger movements of 13 participants while grasping objects and performing everyday tasks (see image below). We recorded their finger movements with an Optitrack system and asked them to perform SoloFinger gestures as well as keep their hand idle. We identify two major thresholds needed to avoid false activations and enable to recognize such gestures: a threshold for idle fingers and a threshold for moving fingers. Our initial results show that for 36 hand-object actions, 23 do not produce false activations, and only 6 produce a significant amount (more details in the paper).

Proof-of-concept: Black-Box Classifier and VR Glove We use a random forest classifier provided by the Sci-kit Learn Python library, and the Hi5 VR glove (see image below). From our initial analysis, we learned that finger movements vary on different objects. Therefore, we performed evaluation in two conditions - training with and without action information. Thus, the classification task was to classify individual trials of all 9 classes (7 gestures + 1 static hold + 1 action without gesture) without action information for every participant. For the condition with action information, we trained models separately for all five actions. With action information, we do not observe false activations and the average accuracy is 89% between all gestures. Without action information, we observe false activations only in 2.12% of the trials (17 out of 800), and an average classification accuracy of 86%.