AT-USTC: A Polysemous USTC Research Designation

Updated 12 July 2026

AT-USTC is a polysemous USTC designation representing distinct research activities, with its clearest reference as the Any-Time Person Re-Identification benchmark.
In the ReID context, the benchmark comprises 403,599 images of 270 identities across six scenarios, enabling robust evaluation across modality and temporal shifts.
Beyond re-identification, AT-USTC identifies USTC systems in translation, meeting diarization, and detector instrumentation, underscoring its cross-domain research significance.

AT-USTC is a context-dependent designation associated with several research activities at the University of Science and Technology of China. In the cited literature, the expression is used explicitly for the Any-Time Person Re-Identification benchmark dataset, and it is also described as a likely or project-specific shorthand in simultaneous translation, meeting diarization, Antarctic telescope development, and timing-detector instrumentation. The explicit usage with the clearest standalone referent is the ReID dataset introduced for Any-Time Person Re-Identification (AT-ReID); elsewhere, the term functions more as an institutional tag than as the formal title of a method or instrument (Li et al., 16 Apr 2026).

1. Terminological scope

The term does not denote a single USTC project across all cited sources. Instead, the literature associates it with several domain-specific referents, some explicit and some interpretive.

Context	Interpretation in source	Formal name in paper
Person re-identification	Explicit benchmark name	AT-USTC dataset
Simultaneous translation	Most likely “Attention/Transducer – USTC”	CAAT
Speech translation pipeline	Label like “AT-USTC” or similar internal system name	USTC-NEL system
Meeting diarization	Referred to as AT-USTC on the leaderboard	USTC-Ximalaya system
Antarctic astronomy	Plausible shorthand for a USTC-associated TianMu unit	Antarctic TianMu / AT-Proto

In the ReID literature, AT-USTC is unambiguously the dataset name. In the simultaneous translation paper, the string “AT-USTC” is said to “almost certainly” refer to an Attention/Transducer–style model developed at USTC, namely the Cross Attention Augmented Transducer (CAAT). In the M2MeT diarization paper, the USTC-Ximalaya system is said to be “often referred to as AT-USTC on the M2MeT leaderboard.” In the Antarctic TianMu telescope paper, the term is not explicit in the paper itself, but the technical summary states that “AT-USTC” can be understood as an Antarctic TianMu telescope associated with USTC (Li et al., 16 Apr 2026, Liu et al., 2021, He et al., 2022, Zhou et al., 30 Dec 2025).

2. AT-USTC as the Any-Time Person Re-Identification benchmark

The most explicit and technically developed meaning of AT-USTC is the dataset used for Any-Time Person Re-Identification (AT-ReID). AT-ReID is defined as retrieval across cloth-change, from short-term to long-term temporal intervals, and under diverse illumination modality including both daytime RGB and nighttime IR. The dataset is described as containing 40,3599 images (199,803 RGB and 203,796 IR) of 270 identities and 710 sets of different clothing, captured from 16 different cameras, specifically 8 RGB and 8 IR devices distributed across 16 non-overlapping locations, covering 5 indoor and 11 outdoor scenes. The accompanying technical summary further notes that the image count formatting implies approximately 403,599 total images (Li et al., 16 Apr 2026).

The benchmark is split into two disjoint identity partitions of 135 identities each. The training set contains 286,087 images, of which 20% (55,060 images) are allocated for validation, while the test set contains the remaining 135 IDs with 117,512 images. The benchmark is organized around six scenarios: DT-ST, DT-LT, NT-ST, NT-LT, AD-ST, and AD-LT, corresponding to daytime or nighttime, short-term or long-term, and all-day mixed-modality settings. The paper defines Any-Time as the average performance of all six scenarios, making robustness across all scenario types the central metric rather than a single-modality or single-timescale score (Li et al., 16 Apr 2026).

This construction differentiates AT-USTC from conventional ReID datasets. The paper contrasts it with traditional ReID datasets such as Market1501 and CUHK03, cloth-changing ReID datasets such as PRCC and LTCC, and cross-modality ReID datasets such as SYSU-MM01. The stated motivation is that methods specialized for only one of these settings encounter a significant performance bottleneck on AT-USTC because they are incapable of addressing retrieval “at anytime,” namely across modality shifts and clothing changes simultaneously (Li et al., 16 Apr 2026).

3. STFER and the benchmark’s operational role

The paper “Beyond Visual Cues: Semantic-Driven Token Filtering and Expert Routing for Anytime Person ReID” introduces STFER, a framework explicitly designed around AT-USTC’s six-scenario structure. STFER uses Qwen3-VL-4B to generate identity-intrinsic semantic text, a ViT-Base/16 backbone, Semantic-driven Visual Token Filtering (SVTF), and Semantic-driven Expert Routing (SER). The model uses 6 scenario-specific CLS tokens, one for each AT-USTC scenario, and the Any-Time objective aggregates scenario-specific identity losses (Li et al., 16 Apr 2026).

On AT-USTC, the reported Any-Time result for Uni-AT is R1 = 55.80, mAP = 41.38, whereas STFER reaches R1 = 94.54, mAP = 93.46. The paper further states that STFER “consistently better results in all six scenarios,” with especially large improvements in long-term and all-day settings (Li et al., 16 Apr 2026).

Scenario	Uni-AT R1 / mAP	STFER R1 / mAP
DT-ST	97.76 / 87.97	98.04 / 95.72
DT-LT	36.75 / 25.89	91.89 / 90.85
NT-ST	81.32 / 53.82	95.12 / 93.52
NT-LT	39.54 / 26.93	98.00 / 97.36
AT-ST	50.25 / 34.94	89.73 / 89.00
AT-LT	29.21 / 18.71	94.46 / 94.29

The ablation study attributes most of the gain to text priors. A baseline without text, SVTF, or SER reproduces the Any-Time score of 55.80 / 41.38; adding text priors raises this to 93.45 / 92.65; adding SVTF yields 93.90 / 92.98; adding SER yields 93.79 / 92.80; and the full STFER reaches 94.54 / 93.46. The model is trained on AT-USTC for 120 epochs on a single NVIDIA RTX 3080 (16 GB), and the paper also reports cross-dataset evaluation on Market1501, CUHK03, SYSU-MM01, PRCC, and LTCC, where a model trained on AT-USTC achieves an average R1 = 74.33, mAP = 75.26 (Li et al., 16 Apr 2026).

In this usage, AT-USTC is not merely an institutional tag; it is the benchmark that defines the AT-ReID problem and the evaluation target for scenario-aware retrieval.

4. Speech and language-system usages

In speech and translation research, AT-USTC is described as a likely shorthand rather than the formal name of the method. In “The USTC-NELSLIP Systems for Simultaneous Speech Translation Task at IWSLT 2021”, the technical summary states that the string “AT-USTC” most likely refers to an Attention/Transducer–style model developed at USTC, namely CAAT (Cross Attention Augmented Transducer). CAAT extends conventional RNN-T to sequence-to-sequence tasks without monotonic constraints, and experiments on simultaneous speech-to-text and text-to-text translation show better quality-latency trade-offs than wait-k. The summary reports an average improvement of 11.3 BLEU for the speech-to-text system and 4.6 BLEU for the text-to-text system over the previous year’s optimal systems (Liu et al., 2021).

A related but earlier use appears in “The USTC-NEL Speech Translation system at IWSLT 2018”. There, the technical summary states that integrated ASR+MT systems from the USTC/IFLYTEK group “often appears under labels like ‘AT-USTC’ or similar internal system names.” The formal system described in the paper is a conventional pipeline with speech recognition, post-processing / re-segmentation, and machine translation. The paper reports that, compared to the baseline system from KIT, the USTC system achieved 14.9 BLEU improvement (Liu et al., 2018).

These two papers indicate that, in speech research, AT-USTC is best understood as an institution-tagged system label rather than the canonical method name. The formal model names are CAAT in simultaneous translation and USTC-NEL in the 2018 pipeline system, but the shorthand emphasizes a USTC Attention/Transducer or ASR-Translation identity (Liu et al., 2021, Liu et al., 2018).

5. System, instrument, and detector usages

Outside language technology, the term is attached to concrete systems and hardware contexts. In the meeting-transcription literature, “The USTC-Ximalaya system for the ICASSP 2022 multi-channel multi-party meeting transcription challenge” states that the USTC-Ximalaya entry is often referred to as AT-USTC on the M2MeT leaderboard. The formal system is a TS-VAD–centric diarization pipeline for multi-channel, heavily overlapped Mandarin meetings. Its final TS-VAD system reduces diarization error rate relative to improved classical clustering by 66.55% on Eval and 60.59% on Test under the reported 0-collar setting (He et al., 2022).

In Antarctic instrumentation, “Antarctic TianMu Staring Observation Project I: Overview and Implementation of the Prototype Telescope” does not explicitly use the string in the paper, but the technical summary states that AT-USTC can be understood as an Antarctic TianMu small-telescope project in which USTC is one of the participating institutions. The prototype telescope, AT-Proto, has an aperture of 18 cm, was transported to Zhongshan Station in October 2022, and “has since operated stably and reliably in the frigid environment for over two years” (Zhou et al., 30 Dec 2025).

Detector-development summaries use the term more loosely, but still in a recognizable project-level sense. The summary for “Performance of USTC first batch resistive AC-LGAD sensor” frames the results “in terms of what is relevant for a timing detector such as an AT-USTC system.” The paper reports spatial resolution can reach 4 μm and a temporal resolution of 48 ps is achieved for the first batch of resistive AC-LGAD sensors (Li et al., 15 Dec 2025). In a related USTC detector program, “Development of Small-pitch, Ultra-thin 3D Silicon Sensors at USTC” reports 50 μm active thickness, 50 μm × 50 μm and 25 μm × 25 μm pixel sizes, and timing results of 39.9 ± 3.6 ps lower-limit resolution for the 50 μm pitch sensor at 110 V and 31.1 ± 2.1 ps for the 25 μm pitch sensor at 96 V (Ma et al., 13 May 2026).

A plausible implication is that, in these hardware contexts, AT-USTC functions as a project-level USTC detector or telescope label rather than as the formal device name. The formal names remain AT-Proto, AC-LGAD, and small-pitch, ultra-thin 3D silicon sensors (Zhou et al., 30 Dec 2025, Li et al., 15 Dec 2025, Ma et al., 13 May 2026).

6. Disambiguation and editorial synthesis

Across the cited sources, AT-USTC behaves less like a single canonical acronym than like a domain-dependent USTC tag. The explicit, self-contained referent is the AT-USTC dataset for Any-Time Person Re-Identification, which defines a six-scenario benchmark over RGB/IR modality changes and short-/long-term clothing variation. In simultaneous translation and speech translation, the same string is described as a likely shorthand for institution-tagged systems, especially CAAT and related USTC ASR + MT pipelines. In meeting diarization, it is a leaderboard name for the USTC-Ximalaya system. In astronomy and detector R&D, it is used or plausibly interpreted as a USTC-associated project label surrounding Antarctic TianMu and USTC timing-detector programs (Li et al., 16 Apr 2026, Liu et al., 2021, He et al., 2022, Zhou et al., 30 Dec 2025, Li et al., 15 Dec 2025).

This suggests that disambiguation depends on the surrounding technical vocabulary. If the context includes Any-Time, ReID, RGB/IR, or STFER, AT-USTC denotes the ReID benchmark. If it includes CAAT, RNN-T, wait-k, or IWSLT, it denotes a USTC Attention/Transducer translation system. If it includes TS-VAD, M2MeT, or ALIMEETING, it denotes the USTC-Ximalaya diarization entry. If it includes AT-Proto, Zhongshan Station, AC-LGAD, or 4D tracking, it denotes a USTC-associated instrument or detector context rather than a single standardized name (Li et al., 16 Apr 2026, Liu et al., 2021, He et al., 2022, Zhou et al., 30 Dec 2025, Li et al., 15 Dec 2025).

In encyclopedia terms, AT-USTC is therefore best treated as a polysemous institutional designation whose meaning is fixed by domain: explicit as a benchmark in Any-Time Person Re-Identification, and otherwise as a USTC system tag spanning language technology, meeting diarization, Antarctic instrumentation, and timing-detector development.