Korea Advanced Institute of Science and Technology (KAIST), M.S. in Computer Science

Thesis

Fishing Things — Fishing: Minimizing False Selections in a Dense Smart Home Environment with Bidirectional Temporal Interaction. Defended 2022-12-19 (committee: Geehyuk Lee, chair; Woohun Lee; In-Young Ko); degree conferred 2023-02. Keywords: Internet of Things, smart home, remote selection, gesture detection, temporal interaction.

In a smart home with many objects, unintended selections grow with the number of objects, and existing remote-selection methods need precise body tracking or a spatial cursor, neither practical at home. The thesis proposes Fishing, a bidirectional time-window selection protocol: the user points at an object, the object shows a visual cue, and the user confirms with a swipe-up gesture timed to that cue. The name comes from angling, where the strike must land while the fish bites. When the receptive fields of several objects overlap, each object opens its time window offset by ΔT so that only one object can be selected at a time.

A central agent tracks the fingertip and dominant eye with a Kinect V2 and ray-casts an eye–fingertip line onto the display (EFRC); a WPF simulation renders eight smart objects (two lights, two speakers, an air conditioner, a projector, a coffee machine, a door lock) across two displays. Receptive-field radius is 11.33 cm (matching the 17.7° pointing error reported in related work). Two preliminary studies fixed the timing parameters: pointing threshold T_PG = 0.4 s, time window TW = 0.7 s (error ≈30% at 0.6 s vs ≈13% at 0.7 s), inter-object cue offset ΔT = 0.6 s. Against dwell selection, Fishing was more accurate (87.7% vs 81.1%; F1 83.5% vs 44.0%) and faster per correct selection (2.33 s vs 3.53 s).

The main study had 16 participants (12 men, 4 women, aged 20–27, about 160 min) and compared Fishing with Slider, a state-of-the-art disambiguation protocol.

• Experiment 1 (unintended gestures): 2.4× fewer false selections from recorded everyday indoor hand movements; accuracy 94.4% vs 89.0%; correct-selection time 2.40 s vs 3.55 s (1.48× faster).
• Experiment 2 (density): at marker gaps of 32.7, 20.8, 14.9 and 11.9 cm, Fishing scored 98.8, 95.1, 93.8 and 95.4% while Slider fell to 98.3, 88.2, 74.2 and 66.8%, a 29.6-point gap at the densest setting. NASA-TLX 37.8 vs 61.0, significantly lower on five of six subscales (all but temporal demand). At gaps above 20 cm, Fishing matched the Kopper natural-pointing model (about 2.1 s).
• Interview: 13 of 16 preferred Fishing; 14 volunteered that it was easy (12), accurate (6) or fast (2), while 15 called Slider inaccurate (10), tiring (9) or complicated (6).

The discussion extends the protocol to receptive-field sizing, Fishing without pointing, alternative confirmation gestures such as voice, and tuning ΔT. The committee credited the topic's importance and the method's originality and validity, and asked for stronger justification of the baseline, of the timing parameters, and of where Fishing sits in the wider interaction once an object is selected.

As first author I led protocol design, implementation, the studies, analysis, and writing. Authors: Auejin Ham, Taejun Kim, Geehyuk Lee.

The studies ran under KAIST IRB approval KH2021-170; I handled the submission through to approval.

Lab

At the HCI Lab (advisor Geehyuk Lee) the research topics were gaze-based menu selection and touch interaction on the mouse wheel. The work involved driving haptic actuators directly, implementing the wireless messaging pipeline, and building the tools used to author haptic interaction. It appeared at ACM CHI 2022 and in the International Journal of Human-Computer Interaction.