TF-PRVR: Training-Free Partially Relevant Video Retrieval

Giyeol Kim; Chanho Eom

cs.CVOct 6, 2026

TF-PRVR: Training-Free Partially Relevant Video Retrieval

Authors: Giyeol Kim, Chanho Eom

Organizations: Chung-Ang University

Abstract

Partially Relevant Video Retrieval (PRVR) aims to retrieve untrimmed videos containing moments relevant to a given text query. Despite recent progress, existing PRVR methods suffer from two key limitations: a fixed video decomposition scheme that causes semantic dilution, and source-domain overfitting induced by task-specific training. In this paper, we propose TF-PRVR, the first training-free framework for PRVR. TF-PRVR leverages frozen vision-language features to construct video-specific hierarchical representations. It derives temporal semantic signals from frame-level features and applies frequency-based multi-scale analysis to identify adaptive temporal boundaries, producing hierarchical segments with coherent event-level semantics. Built on these segments, TF-PRVR constructs a unified multi-scale graph and propagates query relevance across temporally and semantically related nodes. A moment-aware scoring strategy then aggregates temporally aligned relevance across scales, emphasizing consistently supported moments while suppressing isolated false responses. Without task-specific training, TF-PRVR preserves the general-purpose alignment capability of pre-trained vision-language models and avoids dataset-specific overfitting. Extensive experiments demonstrate consistent performance across datasets with diverse visual and temporal characteristics, suggesting a practical direction for training-free PRVR.

Figures & tables

Figure 1 : (a) Dual-branch framework of existing PRVR methods Dong et al. (2022) ; Moon et al. (2025b) ; Wang et al. (2024b) ; Li et al. (2025) ; Moon et al. (2025a) . (b) Semantic dilution caused by fixed-size clip aggregation. (c) Cross-dataset comparison (R@1 (%)) on three datasets (TVR Lei et al. (2020) , ActivityNet Captions (Act) Krishna et al. (2017) , Charades-STA (Cha) Gao et al. (2017) ) using CLIP-L Radford et al. (2021) features. Each group shows in-domain ( e.g. , Act $\rightarrow$ Act) and cross-domain ( e.g. , TVR $\rightarrow$ Act) results for existing methods Wang et al. (2024b) ; Moon et al. (2025b) , zero-shot CLIP, and our training-free TF-PRVR. (Best viewed in color.)

Figure 4 : (a) Overview of TF-PRVR. (b) Video-specific hierarchical segmentation derives temporal semantic signals from frame features and decomposes them into multi-scale boundaries to construct coherent video segments. (c) Graph-based retrieval propagates query relevance over a unified multi-scale segment graph and aggregates scale-consistent responses for moment-aware scoring.

Source Training Data	Method	R@1	R@5	R@10	R@100	SumR
Training-based methods with CLIP-L Radford et al. (2021)
Charades-STA Gao et al. (2017)	MS-SL Dong et al. (2022)	0.4	2.2	3.9	17.3	23.8
	GMMFormer Wang et al. (2024b)	0.4	2.6	4.3	18.8	26.1
	GMMFormerv2 Wang et al. (2024a)	0.6	2.8	4.6	19.6	27.6
	MSC-PRVR Moon et al. (2025b)	0.7	2.8	4.5	20.1	28.1
TVR Lei et al. (2020)	MS-SL Dong et al. (2022)	6.9	19.8	26.5	59.1	112.3

Table 5 : Comparison on ActivityNet Captions Krishna et al. (2017) under cross-dataset transfer and training-free settings. Source training data indicates the PRVR training dataset(s). TF-PRVR † and TF-PRVR ‡ use CLIP-L/14 and EVA-CLIP-L/14 Sun et al. (2023) , respectively, without source-domain PRVR training.

Method	ActivityNet Captions Krishna et al. (2017)					Charades-STA Gao et al. (2017)					TVR Lei et al. (2020)
Method	R@1	R@5	R@10	R@100	SumR	R@1	R@5	R@10	R@100	SumR	R@1	R@5	R@10	R@100	SumR
Training-based methods with ResNet152 He et al. (2016) + I3D Carreira and Zisserman (2017) + RoBERTa Liu et al. (2019)
MS-SL Dong et al. (2022)	7.1	22.5	34.7	75.8	140.1	1.8	7.1	11.8	47.7	68.4	13.5	32.1	43.4	83.4	172.4
PEAN Jiang et al. (2023)	7.4	23.0	35.5	75.9	141.8	2.7	8.1	13.5	50.3	74.7	13.5	32.8	44.1	83.9	174.2
LH Fang et al. (2024)	7.4	23.5	35.8	75.8	142.4	2.1	7.5	12.9	50.1	72.7	13.2	33.2	44.4	85.5	176.3
BGM-Net Yin et al. (2024)	7.2	23.8	36.0	76.9	143.9	1.9	7.4	12.2	50.1	71.6	14.1	34.7	45.9	85.2	179.9

Table 6 : Comparison with state-of-the-art PRVR methods on ActivityNet Captions Krishna et al. (2017) , Charades-STA Gao et al. (2017) , and TVR Lei et al. (2020) . Training-based methods are evaluated in-domain, while training-free methods require no PRVR-specific training. Methods are grouped by training setting and backbone.

Table 7 : Ablation studies on ActivityNet Captions Krishna et al. (2017) . HS, GP, and MAS denote hierarchical segmentation, graph-based relevance propagation, and moment-aware scoring, respectively.

Figure 5 : Qualitative visualization of video-specific hierarchical segmentation on Charades-STA Gao et al. (2017) . The rows show segments at different DWT levels, while the temporal signal is shown below. Red dashed lines connect example queries to their best-matched segments.

Dataset	Method	Source Training Data	R@1	R@5	R@10	R@100	SumR
QVHighlights Lei et al. (2021)	GMMFormerv2 Wang et al. (2024a)	Act+Cha+TVR	18.6	26.5	29.8	50.6	125.5
	MSC-PRVR Moon et al. (2025b)	Act+Cha+TVR	20.1	29.2	33.0	49.8	132.1
	CLIP Zero-shot Radford et al. (2021)	None	21.0	41.4	52.3	82.9	197.6
	TF-PRVR (Ours)	None	23.9	47.5	57.6	89.5	218.5
EPIC-Kitchens-100 Damen et al. (2022)	GMMFormerv2 Wang et al. (2024a)	Act+Cha+TVR	2.5	8.1	12.4	21.1	44.1
	MSC-PRVR Moon et al. (2025b)	Act+Cha+TVR	2.8	8.5	13.1	22.9	47.3

Table A1 : Comparison on additional video datasets. Source training data denotes the PRVR dataset(s) used to train each method. TF-PRVR and CLIP zero-shot require no PRVR-specific training.

Source Training Data	Method	R@1	R@5	R@10	R@100	SumR
Training-based methods with CLIP-L Radford et al. (2021)
Charades-STA Gao et al. (2017)	MS-SL Dong et al. (2022)	1.8	7.1	11.8	47.7	68.4
	GMMFormer Wang et al. (2024b)	2.1	7.8	12.5	50.5	72.9
	GMMFormerv2 Wang et al. (2024a)	2.1	7.8	12.5	50.5	72.9
	MSC-PRVR Moon et al. (2025b)	3.2	12.6	20.1	63.8	99.7
ActivityNet Captions Krishna et al. (2017)	MS-SL Dong et al. (2022)	0.8	3.4	6.2	31.7	42.1

Table A2 : Comparison on Charades-STA Gao et al. (2017) under in-domain, cross-dataset transfer, and training-free settings. Source training data indicates the PRVR training dataset(s). TF-PRVR † and TF-PRVR ‡ use CLIP-L/14 and EVA-CLIP-L/14 Sun et al. (2023) , respectively, without source-domain PRVR training.

Source Training Data	Method	R@1	R@5	R@10	R@100	SumR
Training-based methods with CLIP-L Radford et al. (2021)
TVR Lei et al. (2020)	MS-SL Dong et al. (2022)	31.9	57.6	67.7	93.8	251.0
	GMMFormer Wang et al. (2024b)	29.8	54.2	64.6	92.5	241.1
	GMMFormerv2 Wang et al. (2024a)	34.0	59.7	69.8	94.6	258.1
	MSC-PRVR Moon et al. (2025b)	35.1	61.6	71.5	94.9	263.1
ActivityNet Captions Krishna et al. (2017)	MS-SL Dong et al. (2022)	3.8	9.8	14.2	40.7	68.5

Table A3 : Comparison on TVR Lei et al. (2020) under in-domain, cross-dataset transfer, and training-free settings. Source training data indicates the PRVR training dataset(s). TF-PRVR † and TF-PRVR ‡ use CLIP-L/14 and EVA-CLIP-L/14 Sun et al. (2023) , respectively, without source-domain PRVR training.

Dataset	Segmentation	Mean Best tIoU $\uparrow$	R@0.5 $\uparrow$	GT Coverage $\uparrow$	Fragmentation $\downarrow$	Irrelevant Inclusion $\downarrow$	$D_{\mathrm{all}}\downarrow$
ActivityNet Captions Krishna et al. (2017)	Fixed scheme	0.177	7.2%	0.187	12.41	0.082	0.0864
ActivityNet Captions Krishna et al. (2017)	Ours	0.315	24.8%	0.354	5.86	0.041	0.0592
Charades-STA Gao et al. (2017)	Fixed scheme	0.139	0.1%	0.151	9.78	0.057	0.0109
Charades-STA Gao et al. (2017)	Ours	0.278	14.2%	0.291	4.85	0.031	0.0076
TVR Lei et al. (2020)	Fixed scheme	0.348	27.0%	0.367	5.37	0.076	0.0117
TVR Lei et al. (2020)	Ours	0.462	45.3%	0.491	2.84	0.038	0.0082

Table A4 : Direct analysis of segmentation quality and semantic dilution. Higher values are better for Mean Best tIoU, R@0.5, and GT Coverage, while lower values are better for Fragmentation, Irrelevant Inclusion, and feature dispersion

D_{\mathrm{all}}

. GT moments are used only for post-hoc evaluation.

Boundary Strategy	Budget	Act Krishna et al. (2017)	Cha Gao et al. (2017)	TVR Lei et al. (2020)	QVH Lei et al. (2021)
Uniform fixed intervals (conventional)	160	191.8	80.1	175.4	206.1
Adjacent feature dissimilarity	160	196.2	82.3	179.4	212.6
Self-similarity matrix	160	197.0	83.8	180.1	213.9
Velocity-direction change (Ours)	160	200.7	85.1	183.0	218.4
Ours, unconstrained (video-specific)	Adaptive	200.8	85.4	183.6	218.5

Table A5 : Budget-matched comparison of temporal boundary strategies. Act, Cha, and QVH denote ActivityNet Captions Krishna et al. (2017) , Charades-STA Gao et al. (2017) , and QVHighlights Lei et al. (2021) , respectively. All constrained variants use 160 candidates per video, while the unconstrained TF-PRVR uses a video-specific adaptive budget. We report SumR on all datasets.

Method	Segmentation	GP	MAS	Act	Cha	TVR	QVH
Zero-shot	–	–	–	180.4	77.5	172.2	197.6
+ Single-scale mean pooling (32 clips)	Fixed	–	–	183.5	78.1	173.2	199.8
+ Two-branch (clip + frame)	Fixed	–	–	186.6	78.8	174.4	202.1
+ Multi-scale temporal pooling (3 scales)	Fixed	–	–	188.2	79.2	175.0	203.5
+ Sliding-window multi-scale (exhaustive)	Fixed	–	–	189.5	79.5	175.3	204.8
+ Graph propagation only	Fixed	$\checkmark$	–	190.6	80.0	175.8	205.9

Table A6 : Comparison with training-free baselines using the same frozen features. Act, Cha, and QVH denote ActivityNet Captions Krishna et al. (2017) , Charades-STA Gao et al. (2017) , and QVHighlights Lei et al. (2021) , respectively. All values are SumR.

Table A7 : Ablation studies on ActivityNet Captions Krishna et al. (2017) . We analyze the effects of DWT wavelet basis, the number of DWT decomposition levels, graph components, and propagation strength

\beta

Method	ActivityNet Captions Krishna et al. (2017)			Charades-STA Gao et al. (2017)			TVR Lei et al. (2020)
Method	Full	Short	Long	Full	Short	Long	Full	Short	Long
TF-PRVR (Ours)	193.9	193.7	194.0	80.8	80.7	80.8	178.3	178.4	178.1

Table A8 : SumR performance comparison across video length groups on ActivityNet Captions Krishna et al. (2017) , Charades-STA Gao et al. (2017) , and TVR Lei et al. (2020) , using CLIP-L/14 Radford et al. (2021) features. Videos in each benchmark are partitioned into full, short, and long groups according to video duration.

Method	Indexed Query Latency (ms)	End-to-End Time (s)	Indexed Query Memory (GB)	Peak Deployment Memory (GB)
MS-SL Dong et al. (2022)	54.33	48.74	7.57	8.29
GMMFormerv2 Wang et al. (2024a)	9.74	48.31	1.94	2.66
MSC-PRVR Moon et al. (2025b)	9.78	48.36	0.48	1.20
TF-PRVR (Ours)	17.24	49.08	3.80	4.52

Table A9 : Deployment-level efficiency comparison on ActivityNet Captions Krishna et al. (2017) . End-to-end time includes sampled-frame decoding, frozen-feature extraction, index construction, and one text query.

Explore similar work

Concentrate After Imagination: Text-Conditioned Evidence Grounding for Partially Relevant Video Retrieval
Sep 8, 2026Shuaiqi Cheng, Siyu You, Yanbi Wu +3Visual Evidence Imagination
1The Hong Kong University of Science and Technology (Guangzhou) · University of Electronic Science and Technology of China · 3The Hong Kong University of Science and Technology
CurvSpec: Adaptive Multi-Curvature Learning for Partial Relevant Video Retrieval
Sep 29, 2026Zhen Liu, Letian Li, Jinpeng Wang +4Hyperbolic Learning
SIGS, Tsinghua University Shenzhen, China · Harbin Institute of Technology, Shenzhen Shenzhen, China
Keeping the Evidence Chain: Semantic Evidence Allocation for Training-Free Token Pruning in Video Temporal Grounding
Mar 5, 2026Jiaqi Li, Shuntian Zheng, Yixian Shen +4Video Temporal Grounding Temporal Grounding
University of Warwick, Coventry CV4 7AL, United Kingdom · University of Amsterdam, Amsterdam 1012 WX, Netherlands · Amazon AGI, The United States of America

Sep 8, 2026, cs.CVJ/K move · Enter open · S save

Concentrate After Imagination: Text-Conditioned Evidence Grounding for Partially Relevant Video Retrieval

Shuaiqi Cheng, Siyu You, Yanbi Wu, Yuxi Chen+2

1The Hong Kong University of Science and Technology (Guangzhou) · University of Electronic Science and Technology of China · 3The Hong Kong University of Science and Technology

Visual Evidence Imagination

Open paper AskPDF arXiv

Partially Relevant Video Retrieval (PRVR) retrieves untrimmed videos when queries describe only short moments. Although recent methods improve local representations, uncertainty modeling, and global context, final ranking often still trusts the strongest local response; a coincidentally similar fragment can therefore produce an unsupported peak. We identify this failure as the query-agnostic concentration bottleneck and propose TRACE, a score-level evidence verification operator for PRVR. Given a query and global video registers, TRACE activates query-relevant registers, routes their support to frame-level evidence, and smoothly marginalizes alternative query-to-register-to-frame paths before localized temporal selection. Unlike representation-level feature fusion, TRACE uses this evidence only as a query-conditioned residual calibration of the original local score. On ActivityNet Captions, Charades-STA, and TVR, TRACE achieves the best SumR on all three benchmarks and improves the DreamPRVR backbone by 1.2, 1.1, and 1.5 points, respectively. Ablation, routing-corruption, hard-negative, and cross-backbone transfer analyses support the interpretation that the gains arise from query-conditioned evidence verification rather than a generic score offset.

Abstract

Explore similar work

Concentrate After Imagination: Text-Conditioned Evidence Grounding for Partially Relevant Video Retrieval

CurvSpec: Adaptive Multi-Curvature Learning for Partial Relevant Video Retrieval

Keeping the Evidence Chain: Semantic Evidence Allocation for Training-Free Token Pruning in Video Temporal Grounding