Precise Event Spotting aims to localize fast-paced actions or events in videos with high temporal precision, a key task for applications in sports analytics, robotics, and autonomous systems. Existing methods typically process all frames uniformly, overlooking the inherent spatio-temporal redundancy in video data. This leads to redundant computation on non-informative regions while limiting overall efficiency. To remain tractable, they often spatially downsample inputs, losing fine-grained details crucial for precise localization. To address these limitations, we propose AdaSpot, a simple yet effective framework that processes low-resolution videos to extract global task-relevant features while adaptively selecting the most informative region-of-interest in each frame for high-resolution processing. The selection is performed via an unsupervised, task-aware strategy that maintains spatio-temporal consistency across frames and avoids the training instability of learnable alternatives. This design preserves essential fine-grained visual cues with a marginal computational overhead compared to low-resolution-only baselines, while remaining far more efficient than uniform high-resolution processing. Experiments on standard PES benchmarks demonstrate that AdaSpot achieves state-of-the-art performance under strict evaluation metrics (\eg, +3.96 and +2.26 mAP@0 frames on Tennis and FineDiving), while also maintaining strong results under looser metrics. Code is available at: https://github.com/arturxe2/AdaSpot.
Figures & tables
Figure 1: Illustration of standard PES approaches: (a) high-resolution videos incur high computational cost, whereas (b) low-resolution videos reduce cost but lose fine-grained details crucial for precise temporal localization. In contrast, (c) AdaSpot captures global context from low-resolution videos and adaptively applies high-resolution processing to task-relevant regions, preserving fine-grained details efficiently.
Figure 2: Overview of our proposed method, AdaSpot . (a) The framework uses a low-resolution extractor to process low-resolution clips and generate global features Fl and spatial maps Fs . A RoI selector leverages Fs to identify the most relevant region in each frame. The resulting RoI sequence is then processed by a high-resolution extractor to capture fine-grained features, Fh . Fl and Fh are linearly projected, aggregated, and passed through a temporal modeler , before a prediction head produces per-frame classifications. (b) Details of the RoI selector: channel averaging generates saliency maps from Fs , spatio-temporal smoothing reduces noise, and adaptive-scale RoI selection adjusts the RoI size to the saliency spread.
Tennis
FineDiving
FineGym
F3Set
Cost
Model
δ=0 f
1 f
2 f
0 f
1 f
2 f
0 f
1 f
2 f
0 f
1 f
2 f
P(M)
GFLOPs
E2E-Spot 200MF [ 14 ]
69.78 †
97.01 †
97.68 †
25.00 †
69.01 †
86.24 †
17.50 †
53.44 †
63.73 †
–
–
–
4.49
23.13
E2E-Spot 800MF [ 14 ]
70.04 †
97.31 †
97.86 †
19.47 †
65.49 †
83.83 †
17.90 †
55.05 †
66.06 †
–
–
–
12.70
84.93
UGLF [ 42 ]
–
–
–
–
70.00
87.70
–
50.20
67.80
–
–
–
–
–
T-DEED 200MF [ 51 ]
56.00 †
96.91 †
97.84 †
21.33 †
71.07 †
86.87 †
17.32 †
52.99 †
63.69 †
–
–
–
16.42
21.97
T-DEED 800MF [ 51 ]
58.43 †
97.34 †
97.97 †
19.63 †
69.37 †
85.50 †
18.35 †
53.97 †
64.99 †
–
–
–
64.26
86.34
Table 1: Comparison with state-of-the-art methods for PES on Tennis, FineDiving, FineGym, and F3Set datasets. Bold and underlined values indicate the best and second-best results, respectively. The number of parameters (in millions) and GFLOPs for each method are also reported. For AdaSpot, we show results for two variants AdaSpot s and AdaSpot b (using RegNetY-200MF and RegNetY-400MF as feature extractors, respectively), reporting the mean over three random seeds along with the standard deviation ( ± std ). † denotes results obtained by re-running inference with the provided checkpoints using SNMS with a window of two and 50% overlapping for a fair comparison.
SN-BAS
Cost
Model
δ = 0.5 s
1 s
P(M)
GFLOPs
E2E-Spot 200MF [ 14 ]
51.46
55.11
4.49
40.78
E2E-Spot 800MF [ 14 ]
54.49
58.65
12.70
150.02
T-DEED 200MF [ 51 ]
45.43
48.41
12.31
39.58
T-DEED 800MF [ 51 ]
49.39
53.11
46.22
151.31
Santra et al. [ 34 ]
51.07
55.13
6.84
82.51
Table 2: Comparison of AdaSpot with state-of-the-art methods for the ES setting on SN-BAS. Bold and underlined values indicate the best and second-best results. The number of parameters (in millions) and GFLOPs for each method are also reported. Results for E2E-Spot and T-DEED are obtained by integrating their model into our training pipeline. For Santra et al. [34] , we implement our own version of the ASTRM module following the paper specifications due to the lack of publicly available code. For AdaSpot, we show results for two variants, AdaSpot s and AdaSpot b , and report the mean over three random seeds with standard deviation ( ± std ).
Tennis
SN-BAS
Experiment
δ=0 f
1 f
2 f
δ=0.5 s
1 s
AdaSpot
73.30 ± 0.5
96.90 ± 0.1
97.47 ± 0.1
53.02 ± 0.5
56.43 ± 0.3
(a)
Single-branch counterparts
low-res branch only
71.18
96.73
97.42
47.98
51.69
high-res branch only
71.91
96.62
97.26
52.13
55.95
(b)
Padding type
Table 3: Ablation study of AdaSpot components on Tennis and SN-BAS, evaluating the impact of single branches, padding types, smoothing methods, fixed versus adaptive RoI scales, and auxiliary supervision. We report the mean over three random seeds along with the standard deviation ( ± std ).
Figure 3: Comparison of AdaSpot, a single-branch baseline, and redundancy-aware alternatives across multiple spatial resolutions on Tennis (left) and SN-BAS (right). Each point corresponds to a model configuration, with GFLOPs on the x-axis, mAP on the y-axis, and point size indicating the number of parameters. Models closer to the upper-left with smaller markers achieve better accuracy-efficiency trade-offs.
Figure 4: Qualitative visualization of the saliency maps and the corresponding RoIs selected by AdaSpot across all evaluated datasets: Tennis, FineDiving, FineGym, F3Set, and SN-BAS. In FineDiving and FineGym, events revolve around a main athlete, whereas in Tennis, F3Set, and SN-BAS, they revolve around the ball, which is marked with a star for clarity.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Illustration of the taxonomy of methods addressing spatio-temporal redundancy. We categorize approaches into architecture-based and input-based , and indicate whether each method handles spatial redundancy, temporal redundancy, or both.
Figure 6: Qualitative visualization of saliency maps and the resulting selected RoIs when zero-padding is applied on the Tennis dataset. With zero-padding, the RoI selector ends up biased towards the central part of the frames.
Tennis
SN-BAS
Experiment
δ=0 f
1 f
2 f
δ=0.5 s
1 s
AdaSpot
73.30
96.90
97.47
53.02
56.43
(a)
Fusion mechanism
mean
71.26
96.48
97.07
50.55
54.56
product
71.88
96.75
97.37
51.24
54.95
linear
71.36
96.37
96.96
51.30
55.39
Appendix
Table 4: Extended ablation study of AdaSpot components on Tennis and SN-BAS, evaluating the impact of alternative fusion mechanisms, crop sizes, backbone reuse for both low- and high-resolution branches, adaptive RoI aspect ratios, and multiple RoIs per frame.
Figure 7: Sensitivity of AdaSpot performance ( mAP@0f / mAP@0.5s ) to variations in the threshold parameter τ . The blue line denotes the Tennis dataset (left y-axis), while the green line denotes the SN-BAS dataset (right y-axis).
Low & high-res features
High-res features only
Dataset
AF-v2
Uni-AF
AdaSpot
AF-v2
Uni-AF
AdaSpot
Tennis
70.6 ± 1.5
70.2 ± 1.0
73.3 ± 0.5
68.8 ± 0.9
65.9 ± 0.9
71.9 ± 0.4
SN-BAS
49.0 ± 2.0
49.6 ± 1.3
53.0 ± 0.5
23.1 ± 15.1
39.4 ± 3.3
52.1 ± 0.7
Appendix
Table 5: Comparison of instability under the strictest metric for different learnable cropping methods with AdaSpot. Bold indicates best; (mean ± std. across 3 runs).
Figure 8: Qualitative comparison of the RoIs selected by AdaSpot, AdaFocus-v2, and Uni-AdaFocus on the Tennis (left) and SN-BAS (right) datasets. For visualization, we mark the ball position in each frame with a star, as actions in these datsets occur around the ball; thus, relevant RoIs should contain or closely surround it.
PES
ES
Model
P(M)
GFLOPs
P(M)
GFLOPs
E2E-Spot 200MF [ 14 ]
4.49
23.13
4.49
40.78
E2E-Spot 800MF [ 14 ]
12.70
84.93
12.70
150.02
T-DEED 200MF [ 51 ]
16.42
21.97
12.31
39.58
T-DEED 800MF [ 51 ]
64.26
86.34
46.22
151.31
Santra et al. [ 34 ]
6.46
57.84
6.84
82.51
Appendix
Table 6: Efficiency comparison of AdaSpot with state-of-the-art methods in both the PES setting (typically using 224×224 inputs) and the ES setting (using 398×224 inputs). For each configuration, we report the number of parameters (in millions) and the computational cost in GFLOPs.
Tennis
Post-processing
δ=0 f
1 f
2 f
NMS [ 30 ]
ω=1
62.82
96.93
97.61
ω=2
62.45
96.61
97.64
ω=3
62.35
96.11
97.58
ω=4
62.32
95.94
97.47
ω=5
62.29
95.84
97.29
Appendix
Table 7: Post-processing sensitivity analysis on the Tennis dataset. We report results for standard NMS and Soft-NMS using different window sizes ω . Bold and underlined values indicate the best and second-best results.
AP ( δ=0 f)
Event
Nº observations
E2E-Spot
T-DEED
AdaSpot
HR-only
Far-court ball bounce
8150
76.91
59.47
77.20
75.38
Near-court ball bounce
8127
76.79
68.19
78.98
78.91
Far-court swing
7123
53.24
41.52
64.76
58.77
Near-court swing
7044
56.42
48.85
58.83
58.49
Near-court serve
1690
76.79
67.91
79.24
78.67
Appendix
Table 8: Per-class analysis on the Tennis dataset. For each event class, we report the total number of observations and the AP @0 f results for the best-performing versions of E2E-Spot, T-DEED, and AdaSpot, as well as for an AdaSpot variant using high-resolution features only (HR-only). Event classes are sorted in descending order of observations. The best result per class is highlighted in bold, and the second-best is underlined.
AP ( δ=0 f)
Event
Nº observations
E2E-Spot
T-DEED
AdaSpot
Entry
2984
22.51
24.02
26.74
Som(s).Pike
2152
27.21
23.14
27.58
Som(s).Tuck
1071
31.70
21.72
32.23
Twist(s)
803
18.60
16.45
22.51
Appendix
Table 9: Per-class analysis on the FineDiving dataset. For each event class, we report the total number of observations and the AP @0 f results for the best-performing versions of E2E-Spot, T-DEED, and AdaSpot. Event classes are sorted in descending order of observations. The best result per class is highlighted in bold, and the second-best is underlined.
AP ( δ=0 f)
Event
Nº observations
E2E-Spot
T-DEED
AdaSpot
Uneven bars circles start
6612
11.32
10.26
10.28
Uneven bars circles end
6612
20.19
19.89
19.63
Balance beam leap_jump_hop start
4787
17.52
19.72
18.07
Balance beam leap_jump_hop end
4787
10.31
12.64
10.81
Balance beam flight_salto start
4187
19.84
22.72
24.35
Appendix
Table 10: Per-class analysis on the FineGym dataset. For each event class, we report the total number of observations and the AP @0 f results for the best-performing versions of E2E-Spot, T-DEED, and AdaSpot. Event classes are sorted in descending order of observations. The best result per class is highlighted in bold, and the second-best is underlined.
AP ( δ=0.5 s)
Event
Nº observations
E2E-Spot
T-DEED
AdaSpot
HR-only
Pass
4985
85.15
83.44
85.94
85.11
Drive
4300
81.55
77.25
81.77
81.50
High pass
761
79.30
76.33
78.68
75.18
Header
713
68.14
54.27
68.87
62.62
Ball out of play
551
16.97
19.79
23.01
21.15
Appendix
Table 11: Per-class analysis on the SN-BAS dataset. For each event class, we report the total number of observations and the AP @0.5 s results for the best-performing versions of E2E-Spot, T-DEED, and AdaSpot, as well as for an AdaSpot variant using high-resolution features only (HR-only). Event classes are sorted in descending order of observations. The best result per class is highlighted in bold, and the second-best is underlined.
mAP @0 f
mAP @1 f
mAP @2 f
F1 evt
Edit
F 3 ED [ 24 ]
24.8
60.7
64.8
40.3
74.0
AdaSpot s
53.55 ± 1.2
67.76 ± 0.8
68.41 ± 1.0
48.8 ± 1.1
72.6 ± 0.4
AdaSpot b
55.38 ± 0.3
69.37 ± 0.2
69.94 ± 0.2
51.66 ± 0.6
73.66 ± 0.4
Appendix
Table 12: Comparison of AdaSpot with F 3 ED on the F3Set dataset using standard PES mAP metrics, as well as F1 and Edit scores. Results show the mean over three random seeds with the corresponding standard deviation ( ± ) . Bold and underlined values indicate the best and second-best results.
Precise Event Spotting (PES) requires distinguishing visually similar yet semantically distinct adjacent frames, making it fundamentally different from image classification and coarse action recognition. Although self-distillation methods such as DINO have shown strong representation learning ability in images, we find that directly applying them to PES is ineffective: without supervised guidance, subtle but crucial motion cues are often suppressed as noise, leading to representations that are insensitive to precise event boundaries. To address this, we propose Temporal Feature Distillation, a semi-supervised objective that aligns temporally informative backbone features, rather than projection-head outputs, to preserve motion-sensitive and boundary-aware cues for frame-level localization. A supervised warm-up with a ramp-up schedule further stabilizes training by ensuring that meaningful event cues are learned before unlabeled distillation begins. We also introduce Transformer Gate Shift, a multi-scale gated shifting module that injects motion-aware temporal information into Vision Transformers. Experiments on four fine-grained sports benchmarks show consistent improvements over fully supervised and semi-supervised baselines. Under 10% supervision on FSPerf, our method improves mAP by 4.54 points over the strongest competing approach, and with only 80% labeled data, it matches or surpasses the fully supervised 100% baseline on two of the four datasets.
Hao Xu, Xinyu Wei, Sam Wells +1
Deakin University Melbourne, Australia · Champion Data Melbourne, Australia · Paralympics Australia Melbourne, Australia
Spatio-Temporal Video Grounding (STVG) aims to retrieve the visual trajectory of a specific object from a video stream as described by a natural language expression. However, most advanced methods struggle to balance global context modeling with precise boundary localization. Due to the prohibitive computational costs of processing long videos, these approaches typically resort to low-rate temporal downsampling and implicit motion modeling. This inevitably suppresses high-frequency boundary cues and neglects the explicit inter-frame dependencies required for precise boundary delineation. To address these limitations, we present ScanFocus, a novel coarse-to-fine framework that decouples the STVG task into a global spatio-temporal scan and a local boundary focus. Specifically, we utilize a unified vision-language fusion encoder combined with a lightweight Deformable Semantic-Motion Fusion module to efficiently align multimodal features and generate coarse proposals. To recover the suppressed fine-grained details, we introduce the Semantic-Guided Temporal Aggregator (SGTA) in the refinement stage. By densely sampling around coarse boundaries, SGTA explicitly models short-term temporal interactions under semantic guidance, capturing rapid motion changes for precise timestamp regression. Extensive experiments on three widely used benchmarks demonstrate the performance superiority of our proposed method over previous approaches. Code will be released at https://github.com/TenMinutes209/ScanFocus.
Spatio-temporal action detection in videos requires jointly localizing actors in space and identifying action boundaries over time. A common challenge is constructing temporally stable action tubes, as frame-level detectors often suffer from jitter, fragmentation, and imprecise temporal localization. Many recent approaches address this by introducing heavy spatio-temporal transformers or optical-flow-based pipelines, leading to high computational cost and limited scalability. We propose TubeLite, a lightweight framework for spatio-temporal action detection that focuses on stable tube construction and boundary-aware temporal modeling. TubeLite represents each actor as a tube, defined as a sequence of bounding boxes associated with a single actor over time, and explicitly enforces temporal consistency at both the spatial and semantic levels. The method combines low-jitter actor detection, Gaussian-weighted actor feature extraction, efficient short-term temporal propagation, and a boundary-focused temporal prediction head, while avoiding optical flow and large-scale temporal attention. Despite its compact design, TubeLite achieves strong video-level localization performance. It improves [email protected] by 4.5 and 7.1 percentage points over the best compared method on the MultiSports and UCF101-24 datasets, respectively, with substantially fewer parameters and floating-point operations than transformer-based alternatives, demonstrating that effective spatio-temporal action detection can be obtained through principled, lightweight temporal modeling.
Ali Soltaninezhad, Melissa Cote, Alejandro Rico Espinosa +2
Electrical and Computer Engineering, University of Victoria, Victoria, BC, Canada