Depth-guided Geometric Adaptation for Few-shot Image Classification

Zhiqi Miao; Estefanía Talavera; Weichuan Zhang; Jiapan Guo

cs.CVOct 20, 2026

Depth-guided Geometric Adaptation for Few-shot Image Classification

Authors: Zhiqi Miao, Estefanía Talavera, Weichuan Zhang, Jiapan Guo

Organizations: University of Groningen · University of Twente · Shanxi University of Science and Technology

Abstract

Few-shot image classification (FSIC) frequently suffers from visual ambiguity, where background clutter and RGB appearance similarities hinder discriminative feature matching in data-scarce settings. To mitigate this, we propose DAFIC, a geometry-guided framework that incorporates monocular depth into representation learning, episodic adaptation, and dense support-query matching. DAFIC first learns geometry-aware token representations through RGB-D self-supervised pre-training. Subsequently, the Depth-guided Geometric Adaptation (DGA) module uses depth-derived geometry to refine patch representations with geometry-enriched class context while suppressing background interference. Spatial Bidirectional Patch Matching (SBiM) then performs bidirectional dense matching on the refined features, reducing sensitivity to spurious correspondences and background clutter. Extensive experiments on four benchmarks show that DAFIC achieves strong performance against existing methods, with ablation and robustness analyses supporting the effectiveness of coupling geometry-conditioned episodic adaptation with bidirectional matching for few-shot visual recognition.

Figures & tables

Figure 1: Patch-to-class similarity maps on a MiniImageNet [ 57 ] example. In comparison, DAFIC leverages estimated depth priors to produce more target-focused similarities and suppress responses to background distractors (e.g., leaves/branch/fence).

Figure 2: Overview of our proposed DAFIC framework . The depth estimator first predicts a pixel-wise aligned depth map, which is concatenated with the RGB image to form an RGB-D input. The RGB-D pre-trained ViT backbone then extracts token embeddings, where the CLS token encodes geometry-enriched global semantic context. Next, the Depth-guided Geometric Adaptation (DGA) module refines the patch embeddings: Depth-guided Attention (DA) converts depth priors into geometry-salient patch attention (scaled by $\beta$ ); Geometry-aware Adaptation (GA) then integrates the enriched CLS (class) embedding (scaled by $\lambda$ ). Finally, based on the refined patch embeddings, the non-parametric SBiM classifier performs query-to-support bidirectional dense matching, together with an auxiliary global similarity objective, for reliable FSIC.

Method	Backbone	$\approx$ Params	MiniImageNet		TieredImageNet
Method	Backbone	$\approx$ Params	5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot
ProtoNet [ 53 ]	ResNet-12	12.4M	62.29 ±0.33	79.46 ±0.48	68.25 ±0.23	84.01 ±0.56
CAN [ 24 ]	ResNet-12	12.4M	63.85 ±0.48	79.44 ±0.34	69.89 ±0.51	84.23 ±0.37
FEAT [ 68 ]	ResNet-12	12.4M	66.78 ±0.20	82.05 ±0.14	70.80 ±0.23	84.79 ±0.16
RENet [ 26 ]	ResNet-12	12.4M	67.60 ±0.44	82.58 ±0.30	71.61 ±0.51	85.28 ±0.35
SUN [ 14 ]	ViT	12.5M	67.80 ±0.45	83.25 ±0.30	72.99 ±0.50	86.74 ±0.33

Table 1: Results (%) on MiniImageNet and TieredImageNet. The ± shows 95% confidence intervals. The best and second-best results are shown in bold and underlined . A more comprehensive version of the table is shown in the supplementary materials.

Method	Backbone	$\approx$ Params	CIFAR-FS		FC100
Method	Backbone	$\approx$ Params	5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot
ProtoNet [ 53 ]	ResNet-12	12.4M	72.20 ±0.70	83.50 ±0.50	41.54 ±0.76	57.08 ±0.76
RFS [ 55 ]	ResNet-12	12.4M	71.50 ±0.80	86.00 ±0.50	42.60 ±0.70	59.10 ±0.60
SUN [ 14 ]	ViT	12.5M	78.37 ±0.46	88.84 ±0.32	-	-
FewTURE [ 23 ]	Swin-T	29M	77.76 ±0.81	88.90 ±0.59	47.68 ±0.78	63.81 ±0.75
CPEA [ 20 ]	ViT-S	22M	77.82 ±0.66	88.98 ±0.45	47.24 ±0.58	65.02 ±0.60

Table 2: Results (%) on CIFAR-FS and FC100.

Figure 3: Comparison of patch-to-class semantic alignment between baseline and DAFIC. The ”Baseline” and ”DAFIC” columns visualize the alignment after adaptation, where red indicates higher class relevance and blue indicates lower relevance.

Dataset	Shot	CPEA	CPEA(DGA)	CPEA(DGA+SBiM)	DAFIC
MiniImageNet [ 57 ]	1	72.25±0.65	73.26±0.60	74.02±0.46	77.71 ±0.44
MiniImageNet [ 57 ]	5	87.24±0.38	87.82±0.33	88.12±0.27	90.35 ±0.51
TieredImageNet [ 48 ]	1	77.12±0.70	78.30±0.70	78.86±0.51	80.10 ±0.40
TieredImageNet [ 48 ]	5	90.12±0.45	90.65±0.41	90.61±0.32	91.05 ±0.51
CIFAR-FS [ 2 ]	1	78.27±0.66	78.66±0.60	79.05±0.47	84.30 ±0.44
CIFAR-FS [ 2 ]	5	89.30±0.45	89.92±0.41	89.99±0.33	92.53 ±0.29

Table 3: Controlled comparison of the reproduced CPEA baseline and DAFIC on four FSIC benchmarks. CPEA(DGA) and CPEA(DGA+SBiM) progressively incorporate DGA and SBiM into the baseline. Results are mean accuracy (%) over 1,000 meta-test episodes;

\pm

denotes 95% confidence intervals.

Figure 4: Comparison of RGB- and RGB-D-based pre-training on MiniImageNet. The bars show average test accuracies (%) under 5-way 1/5-shot settings, with a simple prototypical classifier.

Components	1-shot	5-shot
Non-pretrain	30.49 $\pm$ 0.35	41.23 $\pm$ 0.38
RGB-pretrain	69.12 $\pm$ 0.49	84.88 $\pm$ 0.31
RGBD-pretrain	73.85 $\pm$ 0.48	87.64 $\pm$ 0.29
+ DA	74.64 $\pm$ 0.47	88.38 $\pm$ 0.28
+ GA	76.69 $\pm$ 0.51	89.45 $\pm$ 0.35
+ SBiM	77.42 $\pm$ 0.49	90.14 $\pm$ 0.41

Table 4: DAFIC components ablation on MiniImageNet (5-way) with mean accuracy (%) averaged over 1,000 meta-test episodes.

Model	Params (M)	FLOPs (G)	Time (ms)	1-shot Acc.	5-shot Acc.
SMKD [ 36 ]	21.67	339.87	78.40	80.08±0.18	90.63±0.13
CPEA [ 20 ]	21.67	9.38	1.43	78.27±0.66	89.30±0.45
DAFIC	21.76	9.51	1.39	84.30±0.44	92.53±0.29

Table 5: Computational complexity and performance comparison. FLOPs and inference latency are reported exclusively for the few-shot model, with latency averaged over 1800 query images on a single NVIDIA V100 GPU. Parameters (Params) correspond to the ViT-Small backbone. Performance is evaluated on CIFAR-FS.

Figure 5: Example visualization of depth corruption strategies used during inference.

Methods	5-way 1-shot	5-way 5-shot
CPEA	72.25±0.65	87.24±0.38
DAFIC	77.71±0.44	90.35±0.51
DAFIC $\Rightarrow$ Object-region occlusions	73.06±0.61	87.51±0.43
DAFIC $\Rightarrow$ Background occlusions	77.10±0.64	89.81±0.36
DAFIC $\Rightarrow$ Low-quality depth map	74.53±0.65	88.10±0.43

Table 6: Inference robustness to corrupted depth on MiniImageNet. The CPEA results are reproduced.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	# Images	Train/Val/Test
MiniImageNet [ 57 ]	$\sim$ 60,000	64/16/20
TieredImageNet [ 48 ]	$\sim$ 779,165	351/97/160
CIFAR-FS [ 2 ]	$\sim$ 60,000	64/16/20
FC100 [ 44 ]	$\sim$ 60,000	60/20/20

Appendix

Table 7: Commonly adopted FSIC benchmark datasets in prior works.

Category	Configuration
Backbone	ViT-Small with $16{\times}16$ patches; 8192-d projection head shared by class and patch tokens (student/teacher).
[1.5pt/2pt] Optimization	AdamW optimizer; cosine LR from $5\times 10^{-4}$ to $1\times 10^{-6}$ ; weight decay 0.04 $\rightarrow$ 0.4; 10-epoch warm-up.
[1.5pt/2pt] Teacher EMA	EMA momentum: 0.996; teacher temperature increases from 0.04 to 0.07 during the first 30 epochs.
[1.5pt/2pt] Augmentation	RGB-D multi-crop strategy: 2 global crops ( $224{\times}224$ , scale [0.25,1.0]) and 10 local crops ( $96{\times}96$ , scale [0.05,0.25]); RGB-depth synchronized crop/flip; per-channel RGB and depth normalization.
[1.5pt/2pt] MIM	Block-wise masking on global crops (student network); masking ratio sampled from [0.0, 0.3] with variance [0.0, 0.2].
[1.5pt/2pt] Computation	Distributed training with DDP; 3 $\times$ A100 (40GB) GPUs; batch size 384; pre-training for 1,600 epochs (800 for TieredImageNet); checkpoint selected at epoch 1,000 (500 for TieredImageNet).

Appendix

Table 8: Summary of the RGB-D Pre-training Configuration

Panel of Figure 6 — Figure 6: Comparison between our RGB-D and RGB-only self-supervised pre-training across four datasets, using six metrics to monitor the training dynamics. RGB-D pre-training generally achieves lower total and class-level losses, converges faster, and yields higher training accuracy than RGB-only pre-training, while its clustering metrics (NMI and F-score) are generally lower. The yellow and red curves denote the training phases before and after the selected checkpoint used for downstream few-shot classification.

Panel of Figure 7 — Figure 7: More visualizations of patch-to-class semantic alignment between baseline CPEA [ 20 ] and DAFIC. Warmer colors (towards red) indicate stronger class relevance, whereas cooler colors (towards blue) denote weaker or non-relevant regions. As shown, the RGB-D ViT backbone alone already yields more class-consistent responses than the baseline. After applying the full DAFIC framework, the patch-to-class alignment is further improved and becomes more precise.

Method	Venue	Backbone	$\approx$ # Params	MiniImageNet		TieredImageNet
Method	Venue	Backbone	$\approx$ # Params	5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot
MatchNet [ 57 ]	NeurIPS’16	ResNet-12	12.4M	65.64 ± 0.20	78.72 ± 0.15	68.50 ± 0.92	80.60 ± 0.71
ProtoNet [ 53 ]	NeurIPS’17	ResNet-12	12.4M	62.29 ± 0.33	79.46 ± 0.48	68.25 ± 0.23	84.01 ± 0.56
MAML [ 17 ]	ICML’17	ResNet-12	12.4M	49.24 ± 0.21	58.05 ± 0.10	67.92 ± 0.17	72.41 ± 0.20
TADAM [ 44 ]	NeurIPS’18	ResNet-12	12.4M	58.50 ± 0.30	76.70 ± 0.30	62.13 ± 0.31	81.92 ± 0.30
LEO [ 52 ]	ICLR’19	WRN-28-10	36.5M	61.76 ± 0.08	77.59 ± 0.12	66.33 ± 0.05	81.44 ± 0.09

Appendix

Table 9: More comprehensive results (%) on MiniImageNet and TieredImageNet. The ± shows 95% confidence intervals. The best and second-best results are shown in bold and underlined .

Method	Venue	Backbone	$\approx$ # Params	CIFAR-FS		FC100
Method	Venue	Backbone	$\approx$ # Params	5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot
ProtoNet [ 53 ]	NeurIPS’17	ResNet-12	12.4M	72.20 ± 0.70	83.50 ± 0.50	41.54 ± 0.76	57.08 ± 0.76
TADAM [ 44 ]	NeurIPS’18	ResNet-12	12.4M	-	-	40.10 ± 0.40	56.10 ± 0.40
MetaOptNet [ 31 ]	CVPR’19	ResNet-12	12.4M	72.80 ± 0.70	84.30 ± 0.50	47.20 ± 0.60	55.50 ± 0.60
RFS [ 55 ]	ECCV’20	ResNet-12	12.4M	71.50 ± 0.80	86.00 ± 0.50	42.60 ± 0.70	59.10 ± 0.60
DeepEMD [ 70 ]	CVPR’20	ResNet-12	12.4M	-	-	46.47 ± 0.78	63.22 ± 0.71

Appendix

Table 10: More comprehensive results (%) on CIFAR-FS and FC100. The ± shows 95% confidence intervals. The best and second-best results are shown in bold and underlined

DA Design	Initialization
DA Design	RGB	RGB-D
Patch-average	77.89 $\pm$ 0.47	79.15 $\pm$ 0.41
MLP-based	78.21 $\pm$ 0.49	79.43 $\pm$ 0.52
CNN-based	78.86 $\pm$ 0.51	80.10 $\pm$ 0.40

Appendix

Table 11: Accuracy (%) comparison between RGB and RGB-D initialization under different DA designs in the DGA module on the TieredImageNet (5-way 1-shot).

Matching Strategy	MiniImageNet		FC100
Matching Strategy	1-shot	5-shot	1-shot	5-shot
SBiM	77.71 $\pm$ 0.44	90.35 $\pm$ 0.51	50.52 $\pm$ 0.46	69.42 $\pm$ 0.44
SBiM (top 50% patches)	77.02 $\pm$ 0.42	89.92 $\pm$ 0.40	50.15 $\pm$ 0.47	68.15 $\pm$ 0.43
SBiM (one-way)	66.05 $\pm$ 0.47	80.17 $\pm$ 0.36	44.61 $\pm$ 0.44	59.41 $\pm$ 0.43
Mutual-NN (MNN)	76.38 $\pm$ 0.40	89.18 $\pm$ 0.29	49.21 $\pm$ 0.55	67.75 $\pm$ 0.40
MLP (CPEA)	76.20 $\pm$ 0.41	88.79 $\pm$ 0.38	48.56 $\pm$ 0.60	67.05 $\pm$ 0.59

Appendix

Table 12: Comparison of different matching strategies for the DAFIC framework on MiniImageNet and FC100 datasets.

$\beta$ Setting	Learnable	1-shot	5-shot
Fixed $\beta=1.0$	No	76.62 $\pm$ 0.44	89.59 $\pm$ 0.51
Fixed $\beta=5.0$	No	77.02 $\pm$ 0.42	89.92 $\pm$ 0.40
Init $\beta=1.0$	Yes	76.68 $\pm$ 0.51	89.66 $\pm$ 0.36
Init $\beta=5.0$	Yes	77.71 $\pm$ 0.44	90.35 $\pm$ 0.51
Init $\beta=10.0$	Yes	77.34 $\pm$ 0.40	90.22 $\pm$ 0.38

Appendix

Table 13: Ablation on the attention-strength parameter

\beta

on MiniImageNet.

$\lambda$	1-shot	5-shot
$1$	76.97 $\pm$ 0.46	89.50 $\pm$ 0.44
$2$	77.71 $\pm$ 0.44	90.35 $\pm$ 0.51
$4$	77.18 $\pm$ 0.44	89.77 $\pm$ 0.43
$8$	76.99 $\pm$ 0.60	89.61 $\pm$ 0.59

Appendix

Table 14: Ablation on the magnitude

\lambda

on MiniImageNet.

$\alpha$	$\gamma$	1-shot	5-shot
1.0	0	77.42 $\pm$ 0.49	90.14 $\pm$ 0.41
1.0	0.5	77.71 $\pm$ 0.41	90.35 $\pm$ 0.51
1.0	1.0	77.60 $\pm$ 0.44	90.24 $\pm$ 0.46

Appendix

Table 15: Ablation on loss weights

\alpha

for local matching and

\gamma

for auxiliary global supervision on MiniImageNet.

FSL Adaptation	MLP	SBiM
None	30.49 / 41.23	32.93 / 48.10
CPEA-style	31.01 / 46.66	29.73 / 44.78
DGA	32.61 / 49.67	34.36 / 52.81

Appendix

Table 16: Effect of DGA and SBiM without RGB-D pre-training on MiniImageNet. Results are 5-way 1-/5-shot accuracies (%).

Dataset	1-shot	5-shot
MiniImageNet	77.42 $\pm$ 0.26	90.29 $\pm$ 0.15
TieredImageNet	80.09 $\pm$ 0.17	90.94 $\pm$ 0.15
CIFAR-FS	84.45 $\pm$ 0.16	92.40 $\pm$ 0.16
FC100	50.36 $\pm$ 0.21	69.37 $\pm$ 0.32

Appendix

Table 17: Training stability of DAFIC across three runs with different random seeds. Results are reported as mean

\pm

standard deviation (%) across runs.

Method	FSIC Latency	Depth Latency	Total Latency
CPEA	1.43	–	1.43
DAFIC (cached)	1.39	–	1.39
DAFIC (on-the-fly)	1.39	72.63	74.02
SMKD	78.40	–	78.40

Appendix

Table 18: Inference latency with and without cached depth maps. Cached depth maps are precomputed offline, whereas on-the-fly inference includes Depth Anything V2-B.

Method	Depth Source	1-shot	5-shot
CPEA	–	72.25 $\pm$ 0.65	87.24 $\pm$ 0.38
DAFIC	Depth Anything V2	77.71 $\pm$ 0.44	90.35 $\pm$ 0.51
DAFIC	DINOv2-S + DPT	76.16 $\pm$ 0.40	89.05 $\pm$ 0.22

Appendix

Table 19: Sensitivity of DAFIC to an alternative weaker depth prior throughout training and evaluation on MiniImageNet.

Figure 8: Qualitative examples of depth maps generated by Depth Anything V2 (second row) and the weaker DINO-v2 ViT/S + DPT (third row) depth estimator.

Method	Backbone	$\approx$ # Params	MiniImageNet		TieredImageNet
Method	Backbone	$\approx$ # Params	5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot
ProtoNet [ 53 ]	ResNet-12	12.4M	62.29 ± 0.33	79.46 ± 0.48	68.25 ± 0.23	84.01 ± 0.56
ProtoNet+Depth	ResNet-12	12.4M	67.56 ± 0.21	84.11 ± 0.14	72.62 ± 0.42	86.88 ± 0.16
DN4 [ 33 ]	ResNet-12	12.4M	56.93 ± 0.62	73.97 ± 0.49	66.25 ± 0.75	82.62 ± 0.51
DN4+Depth	ResNet-12	12.4M	65.27 ± 0.64	82.33 ± 0.40	70.83 ± 0.73	86.67 ± 0.47
CAN [ 24 ]	ResNet-12	12.4M	63.85 ± 0.48	79.44 ± 0.34	69.89 ± 0.51	84.23 ± 0.37

Appendix

Table 20: Results (%) of FSIC methods with RGB-D inputs on MiniImageNet and TieredImageNet. The

\pm

denotes 95% confidence intervals. Adding depth as an extra input channel only marginally increases the input projection parameters relative to the full model size. For FewTURE, which provides both Swin-T and ViT-S variants, we report the ViT-S results for a backbone-matched comparison.

Method	Backbone	$\approx$ # Params	CIFAR-FS		FC100
Method	Backbone	$\approx$ # Params	5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot
ProtoNet [ 53 ]	ResNet-12	12.4M	72.20 ± 0.70	83.50 ± 0.50	41.54 ± 0.76	57.08 ± 0.76
ProtoNet+Depth	ResNet-12	12.4M	76.17 ± 0.66	86.92 ± 0.44	43.77 ± 0.62	60.22 ± 0.74
DN4 [ 33 ]	ResNet-12	12.4M	71.88 ± 0.68	85.16 ± 0.49	39.56 ± 0.56	56.50 ± 0.54
DN4+Depth	ResNet-12	12.4M	76.23 ± 0.68	88.68 ± 0.41	41.47 ± 0.57	59.21 ± 0.55
RFS [ 55 ]	ResNet-12	12.4M	71.50 ± 0.80	86.00 ± 0.50	42.60 ± 0.70	59.10 ± 0.60

Appendix

Table 21: FSIC method with RGB-D input results (%) on CIFAR-FS and FC100. The ± shows 95% confidence intervals.

Method	Backbone	$\approx$ # Params	CUB		StanfordCars
Method	Backbone	$\approx$ # Params	5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot
ProtoNet [ 53 ]	ResNet-12	12.4M	66.29 ± 0.92	82.58 ± 0.58	70.07 ± 0.18	84.21 ± 0.35
ProtoNet+Depth	ResNet-12	12.4M	68.50 ± 0.52	83.40 ± 0.22	73.21 ± 0.11	85.54 ± 0.20
DN4 [ 33 ]	ResNet-12	12.4M	79.01 ± 0.64	90.03 ± 0.35	86.08 ± 0.59	95.35 ± 0.59
DN4+Depth	ResNet-12	12.4M	81.67 ± 0.64	92.18 ± 0.33	87.21 ± 0.59	96.26 ± 0.59
FEAT [ 68 ]	ResNet-12	12.4M	82.14 ± 0.19	92.05 ± 0.38	86.58 ± 0.65	95.13 ± 0.28

Appendix

Table 22: FSIC method with RGB-D input results (%) on CUB and StanfordCars. The ± shows 95% confidence intervals.

Few-Shot Image Classification

Few-shot image classification learns to recognize visual categories from only a small number of labeled examples.

See all 71 papers

Frequency-Enhanced Dual-Subspace Networks for Few-Shot Fine-Grained Image Classification
Apr 16, 2026Meijia Wang, Guochao Wang, Haozhen Chu +4Frequency-Domain Feature Learning Fine-Grained Image Classification
School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, Xi’an 710000, China
MC-RFM: Geometry-Aware Few-Shot Adaptation via Mixed-Curvature Riemannian Flow Matching
May 8, 2026Salim Khazem, Ibrahim Mohamed Serouis, Zakaria EzzahedFlow Matching Vision Foundation Model Adaptation
Talan Research Center, Paris, France
Training-Only Heterogeneous Image-Patch-Text Graph Supervision for Advancing Few-Shot Learning Adapters
Mar 18, 2026Mohammed Rahman Sherif Khan Mohammad, Ardhendu Behera, Sandip Pradhan +2Few-Shot Image Classification VLM Adaptation
Edge Hill University, Ormskirk, Lancashire, L39 4QP

Apr 16, 2026J/K move · Enter open · S save

Frequency-Enhanced Dual-Subspace Networks for Few-Shot Fine-Grained Image Classification

Meijia Wang, Guochao Wang, Haozhen Chu, Bin Yao+3

School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, Xi’an 710000, China

Frequency-Domain Feature Learning Fine-Grained Image Classification Metric Learning Few-Shot Image Classification Fine-Grained Visual Perception

Open paper AskPDF arXiv

Few-shot fine-grained image classification aims to recognize subcategories with high visual similarity using only a limited number of annotated samples. Existing metric learning-based methods typically rely solely on spatial domain features. Confined to this single perspective, models inevitably suffer from inherent texture biases, entangling essential structural details with high-frequency background noise. Furthermore, lacking cross-view geometric constraints, single-view metrics tend to overfit this noise, resulting in structural instability under few-shot conditions. To address these issues, this paper proposes the Frequency-Enhanced Dual-Subspace Network (FEDSNet). Specifically, FEDSNet utilizes the Discrete Cosine Transform (DCT) and a low-pass filtering mechanism to explicitly isolate low-frequency global structural components from spatial features, thereby suppressing background interference. Truncated Singular Value Decomposition (SVD) is employed to construct independent, low-rank linear subspaces for both spatial texture and frequency structural features. An adaptive gating mechanism is designed to dynamically fuse the projection distances from these dual views. This strategy leverages the structural stability of the frequency subspace to prevent the spatial subspace from overfitting to background features. Extensive experiments on four benchmark datasets - CUB-200-2011, Stanford Cars, Stanford Dogs, and FGVC-Aircraft - demonstrate that FEDSNet exhibits excellent classification performance and robustness, achieving highly competitive results compared to existing metric learning algorithms. Complexity analysis further confirms that the proposed network achieves a favorable balance between high accuracy and computational efficiency, providing an effective new paradigm for few-shot fine-grained visual recognition.

Abstract

Few-Shot Image Classification

Explore similar work

Frequency-Enhanced Dual-Subspace Networks for Few-Shot Fine-Grained Image Classification

MC-RFM: Geometry-Aware Few-Shot Adaptation via Mixed-Curvature Riemannian Flow Matching

Training-Only Heterogeneous Image-Patch-Text Graph Supervision for Advancing Few-Shot Learning Adapters