Temporal action localization (TAL) in wearable sensor streams identifies action classes and temporal boundaries, enabling finer-grained activity understanding than conventional action recognition. However, training typically requires costly start--end annotations for every action instance. To reduce this burden, we study point-supervised TAL, where each instance is labeled with only one timestamp and its class. We propose Progressive Sensor Event Expansion (PSEE), which combines semantic activations, sensor-specific transition evidence, and adaptive temporal ownership to recover point-supervised pseudo segments. These segments supervise standard TAL detectors without modifying their inference procedures. Cross-subject experiments on four inertial-sensing benchmarks demonstrate improved pseudo-boundary quality over adapted point-supervised baselines, compatibility with different TAL detectors, and robustness to point sampling. Code is available at https://github.com/joeeeeyin/PSEE.
Figures & tables
Figure 1: Overview of PSEE for point-supervised sensor TAL. Pseudo segments generated from sparse action points supervise a standard TAL detector during training.
WEAR [ 1 ]
WetLab [ 10 ]
RWHAR [ 12 ]
SBHAR [ 9 ]
Method
@0.3
@0.5
@0.7
mAP
@0.3
@0.5
@0.7
mAP
@0.3
@0.5
@0.7
mAP
@0.3
@0.5
@0.7
mAP
OPOT-Adapt [ 14 ]
82.77
72.28
55.81
70.68
35.76
22.49
8.20
21.67
12.34
9.32
5.20
8.98
68.79
42.73
17.36
42.48
SFNet-Adapt [ 8 ]
84.85
76.15
57.92
74.32
18.94
6.85
1.97
9.12
6.28
1.93
1.67
3.04
85.99
47.01
14.57
50.23
PSEE
84.81
75.76
58.83
74.14
46.43
33.44
19.97
33.55
76.82
67.39
47.73
64.68
61.64
52.70
35.17
50.44
Table 1: Comparison with point-supervised baselines using the same SlimSTAD [ 3 ] detector.
Detector
WEAR
WetLab
RWHAR
SBHAR
SlimSTAD [ 3 ]
74.14
33.55
64.68
50.44
ActionFormer [ 17 ]
66.66
28.68
20.45
80.62
TriDet [ 11 ]
66.42
28.21
33.31
81.19
Table 2: Detector generality of PSEE. Results are mAP over tIoU thresholds 0.3:0.1:0.7 .
Figure 2: Pseudo-boundary quality on four sensor datasets, measured by average tIoU between generated pseudo segments and ground-truth boundaries.
Sampling
WEAR
WetLab
RWHAR
SBHAR
Uniform
72.79
49.25
40.98
69.43
Gaussian
74.34
50.23
48.38
71.91
Table 3: Comparison of point sampling strategies for PSEE. Results are reported as average tIoU (%).
Trans.
Own.
WEAR
WetLab
RWHAR
SBHAR
✓
×
74.71
48.37
39.53
59.88
×
×
71.03
49.82
39.36
57.68
✓
✓
74.34
50.23
48.38
71.91
Table 4: Component ablation of PSEE under Gaussian point sampling. Results are average tIoU (%); ✓ and × denote enabled and disabled components.