Camera motion recognition requires separating viewpoint changes caused by camera translation, rotation, or zoom from object motion and scene dynamics. Video-language models often rely on semantic cues and struggle with the geometric evidence needed for this task. Joint-Embedding Predictive Architectures such as V-JEPA offer strong spatio-temporal representations, but our experiments suggest that they underrepresent geometric cues in some regions. Under temporal-reversal supervision, V-JEPA improves overall but degrades on several motion families, whereas a weaker geometry-oriented VGGT predictor improves across all families. We refer to these geometrically underrepresented regions as geometric blind spots and propose Fovea-JEPA to identify and address them. Using V-JEPA as the base model and VGGT as a geometry expert, we construct intervention pseudo-labels from their prediction errors to weakly supervise a blind-spot predictor. The predictor produces token-wise response maps, which are aggregated into a clip-level intervention score to guide adaptive fusion of visual and geometric features. This selective intervention introduces complementary geometric information while preserving reliable visual predictions. Experiments on CameraMotionDataset, CameraBench, and ShotBench demonstrate the effectiveness of Fovea-JEPA, with gains of 9.67, 8.67, and 24.38 percentage points over previously reported state-of-the-art baselines, respectively. Code and models will be released upon acceptance.
Figures & tables
Figure 1: Motivation of Fovea-JEPA. Top: Fovea-JEPA estimates a blind-spot map Q and aggregates it into an intervention score q , which indicates whether intervention by a geometry expert (e.g., VGGT) is necessary. A high- q case activates geometric intervention to correct the V-JEPA prediction, whereas a low- q case preserves the original prediction. Bottom left: matched-budget intervention controls. Applying intervention to high- q clips outperforms applying it to low- q or randomly selected clips, while retaining VGGT tokens with the highest q scores outperforms retaining those with the lowest scores or using random masks. Bottom right: Fovea-JEPA compared with state-of-the-art methods across three benchmarks.
Figure 2: Overview of Fovea-JEPA. Our dual-branch framework leverages visual-geometric complementary cues and hierarchical error-guided pseudo supervision to detect geometric blind spots. Subsequently, our framework adaptively retains robust visual semantic predictions or activates geometric compensation to correct visual prediction errors.
CameraMotionDataset
CameraBench
ShotBench
Model
Inst. Acc. ↑
Macro-F1 ↑
Weighted-F1 ↑
mAP ↑
AB-Acc. ↑
CM-Acc. ↑
Qwen3-VL-8B-Instruct ( Bai et al. 2025 )
—
—
—
35.94
73.54
44.83
InternVL2.5-8B ( Chen et al. 2025 )
—
—
—
29.50
59.00
33.60
VGGT camera cue ( Feng et al. 2026 )
73.80
87.00
92.00
—
—
—
Cam-Motion-72B ( Lin et al. 2025 )
—
—
—
62.80
80.60
56.47
CamReasoner-7B ( Wu et al. 2026 )
—
—
—
34.30
78.40
39.70
Table 1: Performance of Fovea-JEPA against state-of-the-art methods on three benchmarks.
Figure 3: Qualitative analysis of the learned intervention decisions. Each column pairs two clips with the same ground-truth camera motion. For high- q cases, q>τ activates the VGGT-guided intervention: the predictions from both V-JEPA 2.1 and VGGT are inexact, whereas Fovea-JEPA recovers the exact label set. For low- q cases, q<τ preserves the V-JEPA pathway, retaining its exact prediction despite an inexact VGGT prediction.
Setting
V-JEPA 2.1
VGGT
w/o reversal
76.71 ± 0.53
71.44 ± 0.25
w/ reversal
80.03 ± 0.61
78.72 ± 0.31
Δ
+3.32
+7.28
Table 2: Ablation of temporal-reversal supervision on exact-match accuracy (%). Results are reported as mean ± sample standard deviation over three runs.
Figure 4: Clip-level intervention-budget analysis. The random curve is the expected same-budget accuracy.
Configuration
Inst.
C → W
Net
Full
83.47
59
+42
w/o classification
81.68
54
+20
w/o action
80.05
0
0
w/o keep
83.14
66
+38
w/o map supervision
82.41
57
+29
Table 3: Ablation of the main training objectives.