Semi-supervised fine-grained visual recognition is highly vulnerable to overconfident pseudo-label errors: visually similar categories frequently produce high-confidence yet incorrect predictions, and consistency regularization then reinforces these errors throughout training. Existing semi-supervised learning (SSL) methods estimate pseudo-label reliability almost entirely from the visual classifier itself---maximum probability, adaptive thresholds, or entropy---signals that remain blind to whether a predicted class is semantically compatible with the visual representation. We propose ReCalMatch, a reliability-calibrated semantic framework for semi-supervised fine-grained recognition. Rather than treating textual semantics as auxiliary supervision, ReCalMatch uses multi-aspect semantic prototypes as calibration evidence for pseudo-label learning. We construct class-conditioned semantic prototypes from class names and domain-specific semantic aspects, and measure a visual--semantic agreement score between each unlabeled embedding and its pseudo-label prototype. This agreement is combined with prediction confidence and entropy into a single reliability weight that down-weights pseudo-labels that are visually confident but semantically inconsistent. A semantic consistency term and a semantic margin regularizer further sharpen prototype separability under limited labels. Extensive experiments on CUB-200-2011, Stanford Dogs, NABirds, and iNaturalist18 show that ReCalMatch consistently improves strong SSL baselines, with the largest gains in low-label regimes where pseudo-label noise is most severe.
Figures & tables
Figure 1: Qualitative comparison on CUB-200-2011 under the 1% labeled-data setting. Compared with a visual-only SSL baseline, ReCalMatch produces more object-aligned, class-discriminative activations and suppresses background-dominant responses.
Figure 2: Overview of ReCalMatch . The visual branch performs standard SSL, while the semantic branch constructs multi-aspect class prototypes from frozen CLIP text features. Cross-aspect agreement and disagreement are combined with classifier statistics to calibrate pseudo-label reliability and the unsupervised consistency loss.
Method
Param.(M)
CUB-200-2011
Stanford Dogs
iNaturalist18
NABirds
1%
10%
25%
1%
10%
25%
1%
10%
25%
1%
10%
25%
Pseudo-Label [ 37 ]
25
32.1
58.5
68.3
29.4
56.2
66.5
34.2
55.0
65.9
30.7
55.3
65.2
FreeMatch [ 58 ]
26
35.6
63.4
72.1
33.2
60.5
70.0
36.9
59.4
70.2
32.8
58.7
68.9
FixMatch [ 48 ]
25
40.9
65.7
74.2
36.5
62.8
72.1
39.3
61.1
72.8
35.8
61.9
71.4
FlexMatch [ 65 ]
85
44.5
71.3
80.5
42.2
69.1
78.3
45.2
67.8
78.4
41.6
67.8
77.2
SelfMatch [ 33 ]
80
46.3
73.1
81.3
44.8
71.5
80.2
47.1
69.5
79.9
43.7
69.4
78.8
Table 1: Performance of SSL methods on fine-grained datasets. Top-1 accuracy (%) under different labeled-data settings. The bold and underlined figures represent the best and second-best results, respectively.
Figure 3: t-SNE visualization of feature embeddings produced by different methods on CUB-200-2011. Left to right: visual-only SSL baseline (FreeMatch), SemiVisBooster with class-level text guidance, and ReCalMatch with reliability-calibrated semantic guidance.
Figure 4: Qualitative comparison of discriminative region localization across models with different semantic supervision settings. Left to right: input image, visual-only SSL baseline (FreeMatch), SemiVisBooster with class-level text guidance, and ReCalMatch with reliability-calibrated semantic guidance.
Configuration
1%
10%
25%
FreeMatch
35.6
63.4
72.1
+ Lscon (class-name)
50.9
74.8
80.5
+ multi-aspect prototypes
52.3
77.3
81.2
+ Lmargin
53.0
77.6
81.6
+ reliability calibration
53.5
77.9
81.9
+ disagreement gating (full)
54.2
78.2
82.3
Table 2: Component ablation (Top-1 accuracy – %) on CUB-200-2011 under different labeled-data settings.
Figure 7
Group
Setting
1%
10%
25%
(λs,λm)
(0.1, 0.05)
51.9
76.6
81.4
(0.5, 0.05)
53.4
77.8
82.0
(1.0, 0.05)
54.2
78.2
82.3
(2.0, 0.05)
53.7
77.9
82.1
(1.0, 0.10)
54.0
78.1
82.2
(α,β)
(0.0, 0.0)
51.5
76.8
81.1
Table 3: Compact sensitivity and stability analysis on CUB-200-2011 – Top-1 accuracy (%). Multi-seed rows report mean ± std over three runs.
Semantic-bank source
Top-1 (%)
Class names only
50.9
LLM-generated aspects
53.6
Curated metadata aspects
54.2
LLM-generated + metadata
54.8
Table 4: Effect of semantic-bank source on CUB-200-2011 under the 1% labeled-data setting.