---
title: "Adelaide's HiCE Spots the Exact Hand-Touch Frame With a 16.91 Precision Gain"
subtitle: "A new hand-centric attention module and a 4,021-video dataset push frame-accurate touch detection well beyond existing event-spotting baselines."
url: "https://alphasignal.ai/news/adelaide-s-hice-spots-the-exact-hand-touch-frame-with-a-16-91-precision-gain"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-04-14T00:00:00.000Z"
updated: "2026-08-24T00:24:42.362Z"
type: "Paper"
topics:
  - "Robotics"
  - "Image"
subtopics:
  - "Vla Models"
  - "Manipulation"
  - "Pose Estimation"
upvotes: 2635
source_url: "https://arxiv.org/abs/2604.12343"
comments: 0
truncated: true
---

# Adelaide's HiCE Spots the Exact Hand-Touch Frame With a 16.91 Precision Gain

> A new hand-centric attention module and a 4,021-video dataset push frame-accurate touch detection well beyond existing event-spotting baselines.

Detecting the exact frame when a hand first touches an object sounds like a simple perception task, but in first-person video it becomes one of the harder temporal-localization problems in computer vision. Researchers from the Australian Institute for Machine Learning at Adelaide University have proposed [HiCE](https://arxiv.org/abs/2604.12343) (Hand-informed Context Enhanced module), together with a new benchmark called TouchMoment, and report a 16.91 average-precision gain over the strongest prior event-spotting methods.

## Why a single frame is so hard to find

Touch moments mark critical temporal boundaries in manipulation, indicating when an action begins, transitions, or completes, and they carry rich cues for understanding human behavior, assessing motor capability, analyzing skilled performance, and transferring dexterous demonstrations to robots. The catch is that they last exactly one frame. The authors focus on the first touch moment, the single frame at which contact between the hand and an object is first established, treating it as an instantaneous event rather than a temporal segment.

In head-mounted footage, that frame is easy to miss. Rapid head motion introduces strong egomotion, approaching hands often create severe occlusions and perspective distortions at close range, and subtle micro-motions immediately before contact, fine-grained pre-touch hand shaping, and motion blur further obscure the true moment of contact. Prior work on onset detection was built for scenarios that do not share these constraints. Facial-action-unit onset methods rely on static cameras, and sports action-spotting methods operate on videos where the target activity dominates the frame and the subject stays consistently visible.

## Building a benchmark from scratch

To even measure progress on this problem, the team had to build a dataset. TouchMoment comprises 4,021 videos sourced from diverse environments, object categories, and interaction contexts, and includes 8,456 manually annotated touch moments recorded at the frame level. The clips are drawn from two existing corpora: HOI4D, a large-scale benchmark for category-level human-object interaction with egocentric RGB-D sequences, dense 3D hand pose, object pose, and panoptic segmentation, and the egocentric portion of TACO, containing bimanual tool and object manipulation captured with a head-mounted RGB camera at 30 fps, with continuous patterns such as grasping, re-grasping, and coordinated two-hand motion.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [Detecting Precise Hand Touch Moments in Egocentric Video](https://arxiv.org/html/2604.12343)
- [HiCE code repository](https://github.com/bbvisual/hice)
- [T-DEED: Temporal-Discriminability Encoder-Decoder for Precise Event Spotting](https://arxiv.org/abs/2404.05392)