THU EACT-50: Event-Based Action Recognition Benchmark

• THU EACT-50 is a large-scale event-based action recognition benchmark featuring 10,500 recordings of 105 subjects over 50 action classes, captured under varied conditions.
• The dataset supports diverse methodologies, from compact holographic spatiotemporal representations to dense, translation-invariant mappings, demonstrating robust temporal dynamics.
• A lighting-challenging variant (THU-EACT-50-CHL) and a multi-view extension (THU MV-EACT-50) facilitate detailed evaluation of early and fine-grained event-based recognition.

Searching arXiv for papers on THU EACT-50 and related event-based action recognition benchmarks. THU EACT-50 is a large-scale event-based human action recognition benchmark directly recorded with an event camera and introduced by Gao et al. in IEEE TPAMI 2023, as cited by later work. In subsequent arXiv papers, it is described as comprising 10,500 recordings of 105 subjects over 50 action classes, with raw event streams at 1280×800 resolution and durations of 2 to 5 seconds; a commonly used evaluation variant, THU-EACT-50-CHL, contains 2,330 samples across the same 50 categories and is characterized by varying or challenging lighting conditions (Neumeier et al., 10 Jul 2025, Hao, 4 Feb 2026).

1. Origin, naming, and scope

The benchmark is referred to in later papers as THU EACT-50 or THU-EACT-50. The cited texts do not provide a formal expansion of “EACT”; one paper states only that, in context, it denotes an event-camera action dataset from Tsinghua University. The benchmark’s use and protocol are consistently attributed to Gao et al., “Action recognition and benchmark using event cameras,” IEEE TPAMI, 45(12):14081–14097, 2023, which later arXiv works cite as the primary dataset source (Fan et al., 24 Jan 2026).

A related naming point concerns the CHL suffix. Later papers do not provide an official expansion, but they describe THU-EACT-50-CHL as a subset or variant “captured under various lighting conditions” and characterize it as the lighting-challenging split used for evaluation. This establishes CHL as a robustness-oriented benchmark condition rather than a separate action taxonomy (Fan et al., 24 Jan 2026).

The dataset also has a multi-view successor. The paper “Hypergraph-based Multi-View Action Recognition using Event Cameras” introduces THU MV-EACT-50 as an explicit multi-view extension of the earlier single-view dataset, stating that it keeps the same 50 action classes while capturing them simultaneously from multiple synchronized cameras (Gao et al., 2024).

2. Composition and action taxonomy

Later work describes THU EACT-50 as containing 10,500 recordings, 105 subjects, and 50 action classes. The recordings are event streams with durations of 2 to 5 seconds, and EEvAct reports the raw event-stream resolution as 1280 × 800 (Neumeier et al., 10 Jul 2025).

The multi-view extension clarifies the class vocabulary by stating that it preserves the same 50 action classes as the earlier single-view benchmark. Its published list spans several semantic groups: whole-body motions such as walking, running, jump up, running in circles, and squat down; health-related events such as falling down, headache, stomachache, back pain, and vomit; fine-grained motions such as nod head, shake head, thumb up, clap, and wipe face; object interactions such as calling with phone, reading, throw, drink water, open umbrella, close umbrella, put on glasses, take off glasses, put on bag, and take object out of bag; and two-person interactions such as shake hands, fighting, handing objects, and lifting chairs (Gao et al., 2024).

EEvAct further notes that the action inventory includes difficult confusions, explicitly citing put on glasses vs. put off glasses and the triplet sit down, stand up and sit and stand. This indicates that the dataset was used not merely for coarse action categorization but also for fine-grained discrimination under event-based sensing (Neumeier et al., 10 Jul 2025).

3. Event modality and annotation structure

Across later papers, THU EACT-50 is consistently treated as an event-camera dataset whose basic observation is an asynchronous tuple

$e=(x,y,t,p), \quad p \in \{-1,+1\}.$

Here, $x$ and $y$ denote pixel coordinates, $t$ the timestamp, and $p$ the polarity. HoloEv-Net writes the event set as

$$\mathcal{E} = \{e_k\}_{k=1}^N = \{(x_k, y_k, t_k, p_k)\}_{k=1}^N,\quad p_k \in \{-1, 1\},$$

with positive and negative subsets $\mathcal{E}_+$ and $\mathcal{E}_-$ (Hao, 4 Feb 2026).

The multi-view paper also restates the event-camera sensing model in terms of a log-brightness threshold, noting that for a pixel at $(x,y)$, an event is triggered at time $t$ when

$x$0

This is presented as a primer on event sensing rather than a dataset-specific file specification, but it clarifies the modality assumed by downstream methods (Gao et al., 2024).

Later papers give relatively little information about on-disk data formats, annotation schemas beyond per-clip action labels, or licensing. SMV-EAR explicitly states that licensing and download links are not provided in its text and refers readers to the original Gao et al. publication for official access and licensing. HoloEv-Net likewise defers sensor-model, clip-duration, and collection specifics for THU-EACT-50-CHL to the original source (Fan et al., 24 Jan 2026, Hao, 4 Feb 2026).

4. THU-EACT-50-CHL as the lighting-challenging variant

A substantial portion of the recent literature evaluates not on the full THU EACT-50 benchmark but on THU-EACT-50-CHL. HoloEv-Net reports that this variant contains 2,330 samples across 50 categories and describes it as featuring distinct lighting conditions. SMV-EAR describes the same variant as being “captured under various lighting conditions,” emphasizing robustness under illumination changes (Hao, 4 Feb 2026, Fan et al., 24 Jan 2026).

The cited papers do not fully restate the official split structure for CHL. HoloEv-Net says that evaluation follows the dataset’s official protocol but does not enumerate train, validation, and test counts. SMV-EAR similarly follows the established protocol and reports Top-1 accuracy with mean ± standard deviation over five random seeds, but does not specify whether the split is cross-subject, cross-scene, or another partitioning. This means that CHL is operationally well established in later benchmarking, but its exact protocol must still be checked against the primary dataset source (Fan et al., 24 Jan 2026).

The metrics used on CHL vary slightly by paper. HoloEv-Net reports Top-1 and Top-5 accuracy together with parameters, FLOPs, and measured inference latency (ms/sample). SMV-EAR reports Top-1 accuracy, FLOPs, and parameter counts on THU-EACT-50-CHL, while Top-5 is not reported there for this dataset (Hao, 4 Feb 2026, Fan et al., 24 Jan 2026).

A notable descriptive inconsistency appears in the later literature. SMV-EAR reports THU-EACT-50-CHL as using an event camera at 346×260 spatial resolution, whereas HoloEv-Net states that the sensor model and resolution are not detailed in that paper. This suggests that later papers do not all reproduce the same level of dataset metadata for the CHL variant (Fan et al., 24 Jan 2026, Hao, 4 Feb 2026).

5. Common preprocessing and benchmark usage

THU EACT-50 has been used to evaluate several distinct representation paradigms, and those paradigms reveal what later authors consider important properties of the benchmark: fine temporal dynamics, view-dependent motion cues, and robustness to nuisance variation.

HoloEv-Net converts raw events into a Compact Holographic Spatiotemporal Representation (CHSR),

$x$1

with three channels: positive density, negative density, and a holographic map encoding horizontal spatial cues in the $x$2–$x$3 domain. Following MVF-Net, it discretizes the temporal axis to $x$4. The method defines density maps

$x$5

and a holographic map

$x$6

On THU-EACT-50-CHL, HoloEv-Net-Base reports 69.85% Top-1, 79.60% Top-5, 12.5M parameters, 1.9G FLOPs, and 9.6 ms latency; HoloEv-Net-Small reports 67.10% Top-1, 78.86% Top-5, 4.4M parameters, 0.1G FLOPs, and 7.4 ms latency (Hao, 4 Feb 2026).

SMV-EAR instead uses a translation-invariant dense conversion over the temporal views $x$7–$x$8 and $x$9–$y$0. It models events as

$y$1

discretizes the temporal axis to $y$2, and produces two-channel dense maps per view using event count and polarity sum, with both $y$3 and $y$4 resized to 224×224. Its practical conclusion is that global aggregation along the orthogonal axis avoids translation-variant artifacts introduced by spatial binning. With dynamic late fusion and Diverse Temporal Warping (DTW) augmentation, it reports 66.7% Top-1 (±0.29) on THU-EACT-50-CHL at 3.6G FLOPs and 23.5M parameters (Fan et al., 24 Jan 2026).

EEvAct uses THU EACT-50 differently: as a benchmark for early event-based recognition. It center-crops the raw streams to 600×600, downsamples to 100×100, and bins events into frames using a constant time bin, with 2 ms identified as the best setting for early accuracy. Training uses random 1000 ms crops from each sample, whereas evaluation starts from the beginning of each recording and reports Top-1 and Top-5 at growing observation times $y$5. Its best final model, Two-Path adLIF + EGRU, reaches 94.9% Top-1 at 2.0 s, while the paper states that the best early-recognition model attains 50% Top-5 accuracy within 100 ms of observation (Neumeier et al., 10 Jul 2025).

Work	Input protocol on THU EACT-50	Reported result
HoloEv-Net (Hao, 4 Feb 2026)	CHSR, $y$6, 3-channel $y$7–$y$8 tensor	69.85% Top-1 on THU-EACT-50-CHL
SMV-EAR (Fan et al., 24 Jan 2026)	TISM on $y$9–$t$0 and $t$1–$t$2, $t$3, 224×224 maps	66.7% Top-1 on THU-EACT-50-CHL
EEvAct (Neumeier et al., 10 Jul 2025)	2 ms bins, 600×600 crop, 100×100 input, early recognition protocol	94.9% Top-1 on THU EACT-50 at 2.0 s

These divergent preprocessing choices indicate that the benchmark supports both compact spatiotemporal projection methods and high-rate streaming recognition. A plausible implication is that THU EACT-50 is valuable not only as a static classification benchmark but also as a stress test for temporal modeling assumptions.

6. Extension to multi-view benchmarking and reported ambiguities

The most explicit extension of THU EACT-50 is THU MV-EACT-50, introduced as a six-view dataset that keeps the same 50 action classes while expanding the capture setup to synchronized multi-camera recording. The paper reports 31,500 recording sequences, 105 subjects, 6 viewpoints, 1280×800 resolution from CeleX-V event cameras, and an average clip length of 2.34 seconds. Its stated goal is to address multi-view-specific issues such as information deficit and semantic misalignment (Gao et al., 2024).

THU MV-EACT-50 also formalizes benchmark protocols absent from later CHL papers: a cross-subject split with 85/10/10 train/validation/test subjects, and a cross-view split with 4/1/1 training/validation/test views. HyperMV reports 95.74% Top-1 in cross-subject evaluation and 58.54% Top-1 in cross-view evaluation, with 99.23%/99.90% Top-3/Top-5 for cross-subject and 78.07%/83.92% for cross-view. These figures pertain to the multi-view extension rather than the original single-view benchmark, but they clarify how the THU EACT-50 taxonomy has been generalized for viewpoint-rich recognition (Gao et al., 2024).

Several ambiguities remain across the later literature on the original dataset. EEvAct describes THU EACT-50 as 1280×800 and 2 to 5 s long; HoloEv-Net gives no sensor or duration details for THU-EACT-50-CHL; SMV-EAR reports 346×260 for THU-EACT-50-CHL. Likewise, later CHL papers usually defer official splits, licensing, and raw-format details to Gao et al. rather than restating them. This suggests that cross-paper comparisons require care, especially when the comparison spans the full dataset, the CHL variant, and the multi-view extension (Neumeier et al., 10 Jul 2025, Hao, 4 Feb 2026, Fan et al., 24 Jan 2026).

In the current literature, THU EACT-50 therefore functions as a family of closely related benchmarks centered on the same 50-action vocabulary: the original single-view dataset, the lighting-challenging CHL variant, and the six-view MV extension. Their shared role is to provide event-camera action sequences rich enough to evaluate temporal precision, robustness to nuisance conditions, and, in the multi-view case, viewpoint-aware fusion.