Cross-polarization pairing for label-free few-shot thin-section classification

Bingchen Yan

cs.CVOct 17, 2026

Cross-polarization pairing for label-free few-shot thin-section classification

Authors: Bingchen Yan

Organizations: School of Computer Science and Information Security, Guilin University of Electronic Technology, Guilin, 541004, China

Abstract

Few-shot classifiers of rock thin sections are meta-trained on many rock types that a petrographer has identified slide by slide, and this identification is the costly step that new collections lack. We study few-shot recognition of unseen rock types when no training slide is identified, and the only information available is what the microscope records anyway: which slide an image comes from, and whether it was taken in plane-polarized (PPL) or cross-polarized (XPL) light. We first show why generic unsupervised pseudo-tasks fit this setting poorly: the positives they find without labels are either wrong or near-duplicates of the anchor, so they teach the model little. Polarized-light microscopy provides positives that avoid both problems. The PPL and XPL views of one field show the same grains, so they are correct by construction, yet they look very different. We propose cross-polarization pairing (XPP), which treats each slide as a pseudo-class and pairs its PPL and XPL views of the same field location, so that matching must rely on grain shape and texture rather than on colour. On the Nanjing University collection of 108 rock types, with support and query images taken from different slides, XPP is the best label-free method on average over three random splits, and matches a supervised model trained on 78 identified rock types. Controls show that colour augmentation and other acquisition positives do not reproduce the gain. We also show that the image-level protocol of earlier work is saturated, and propose a cross-slide protocol in its place.

Figures & tables

Rock class	Rock types	Slides	Images
Igneous	40	120	963
Metamorphic	40	120	972
Sedimentary	28	84	699
Total	108	324	2634

Table 1: Nanjing University thin-section collection (counts from released archives).

Figure 1: Aligned XPL and PPL frames of one field for an igneous (anorthosite), a metamorphic (metadiabase) and a sedimentary (bauxite) slide, after preprocessing. Dashed box: one XPP anchor box; right columns: two 224-px shots it yields. Interference colours and extinction (XPL) are replaced by relief and absorption colour (PPL), while every grain boundary stays in place.

Positive source (thin sections)	Needs labels?	Precision	Mean cosine
Augmentation of the same crop	no	1	0.931
Feature nearest neighbour, other slides	no	0.348	0.860
XPL at another angle, same location	no	1	0.857
Non-overlapping region, same image	no	1	0.840
XPL at another angle, another region	no	1	0.826
PPL $\leftrightarrow$ XPL, same location	no	1	0.689

Table 2: Candidate positive sources on thin sections, measured with ImageNet ResNet-50 features (323 slides, 108 rock types, 969 pairs per row).

Figure 2: Positive sources on thin sections: mean cosine to anchor (ImageNet ResNet-50). Filled markers have precision 1; vertical line marks oracle (another slide of same rock type). Only PPL view of same box is both correct and less similar than oracle.

Figure 3: XPP framework, drawn with real slides from Nanjing collection (igneous, metamorphic, sedimentary). (a) Each unlabelled slide is one pseudo-class; its shots are one box cropped from aligned XPL frame and from PPL frame of same field. (b) One ImageNet-pretrained ResNet-50 embeds both views; conv1–layer3 are frozen and only layer4 is updated. (c) Shots are pulled to normalized mean of their slide’s support shots by a cosine softmax over 32 slides of each task. (d) At test time, an XPL image is described by mean feature of a 2 $\times$ 3 grid of crops and assigned to nearest prototype built from another slide. Network is standard; contribution lies in how pseudo-tasks are built.

Method	Rock identifications	Cross-slide 1-shot	Cross-slide 5-shot	Image-level 1-shot	Image-level 5-shot
ImageNet features	0	84.96	86.43	89.49	97.31
Instance training	0	87.32 $\pm$ 0.19	88.27 $\pm$ 0.20	91.19 $\pm$ 0.14	97.65 $\pm$ 0.05
Colour-jitter control	0	84.50 $\pm$ 0.18	85.83 $\pm$ 0.26	88.99 $\pm$ 0.24	95.80 $\pm$ 0.09
Neighbour pseudo-tasks	0	87.12 $\pm$ 0.47	88.06 $\pm$ 0.48	91.05 $\pm$ 0.31	97.62 $\pm$ 0.05
XPP (ours)	0	89.32 $\pm$ 0.27	90.48 $\pm$ 0.30	92.62 $\pm$ 0.19	98.04 $\pm$ 0.09
Supervised	78 types	88.85 $\pm$ 0.36	89.79 $\pm$ 0.34	92.24 $\pm$ 0.25	96.83 $\pm$ 0.10

Table 3: Main results on primary split: 5-way accuracy (%), mean

\pm

sd over three seeds, 2000 test episodes.

Figure 4: Cross-slide 5-way 1-shot accuracy on primary split (mean

\pm

sd over three seeds). Diamonds use rock identifications.

Shots drawn from	Cross-slide 1-shot	vs instance training
Same XPL box (instance training)	87.32 $\pm$ 0.19	–
+ other regions of the slide	86.11 $\pm$ 0.31	$-$ 1.21
+ other rotation angles, same box	86.89 $\pm$ 0.24	$-$ 0.42
+ other regions + other angles	85.69 $\pm$ 0.08	$-$ 1.63
+ PPL, same box (XPP)	89.32 $\pm$ 0.27	+2.00
+ PPL + other regions + other angles	87.85 $\pm$ 0.38	+0.53

Table 4: Acquisition positives added to instance training (primary split, cross-slide 5-way 1-shot, %).

Figure 5: Change in cross-slide 1-shot accuracy when acquisition positives are added to instance training (primary split, three seeds).

	Gentle	Aggressive
Method	(layer4, lr 1e-5, 1500 it)	(layers 3–4, lr 1e-4, 4000 it)
Instance training	87.32	76.03
Other regions + angles	85.69	80.42
PPL + regions + angles	87.85	79.50
Neighbour pseudo-tasks	87.12	82.36
Supervised	88.85	89.46

Table 5: Gentle versus aggressive label-free adaptation (primary split, cross-slide 5-way 1-shot, %). Gentle: mean of three seeds; aggressive: one seed.

Figure 6: Gentle (blue) versus aggressive (grey) adaptation. Dashed line marks ImageNet features without training.

Split	ImageNet	Instance	Colour jitter	Neighbour	XPP	Supervised
0 (primary)	84.96	87.32 $\pm$ 0.19	84.50 $\pm$ 0.18	87.12 $\pm$ 0.47	89.32 $\pm$ 0.27	88.85 $\pm$ 0.36
1	85.35	85.39 $\pm$ 0.36	88.89 $\pm$ 0.25	89.56 $\pm$ 0.39	89.87 $\pm$ 0.10	88.57 $\pm$ 0.38
2	81.58	84.36 $\pm$ 0.28	85.34 $\pm$ 0.14	84.29 $\pm$ 0.16	85.04 $\pm$ 0.62	87.19 $\pm$ 0.20
Mean of splits	83.96	85.69	86.24	86.99	88.08	88.20

Table 6: Cross-slide 5-way 1-shot accuracy (%) on three random rock-type splits.

Training signal	Rock identifications	Cross-slide 1-shot	Cross-slide 5-shot
ImageNet features	0	84.96	86.43
Supervised	10 types (30 slides)	85.78 $\pm$ 0.09	87.15 $\pm$ 0.08
Supervised	20 types	86.13 $\pm$ 0.31	87.04 $\pm$ 0.32
Supervised	40 types	87.22 $\pm$ 0.23	88.27 $\pm$ 0.30
Supervised	78 types (all)	88.85 $\pm$ 0.36	89.79 $\pm$ 0.34
XPP	0	89.32 $\pm$ 0.27	90.48 $\pm$ 0.30

Table 7: Label efficiency on primary split: supervised meta-training with

k

identified rock types versus XPP (%).

Figure 7: Supervised meta-training with

k

identified rock types (black) compared with XPP, which uses none (blue line).

	PPL–XPL cosine		PPL retrieves its
Encoder	Same field	Other fields	XPL field (top-1)
ImageNet	0.743	0.636	25.0%
Instance training	0.740	0.615	16.7%
Colour-jitter control	0.699	0.460	33.3%
XPP	0.819	0.651	50.0%
Supervised	0.674	0.491	13.3%

Table 8: Cross-illumination alignment of encoders (primary split, seed 1, 20 unseen test rock types). Retrieval: share of PPL images whose nearest XPL image among all test fields is their own field.

	Cross-slide	Cross-slide 1-shot,	20-way, leave-
Encoder	1-shot	greyscale images	one-slide-out
ImageNet	84.96	83.55	75.95
Instance training	87.23	88.34	79.29
Colour-jitter control	84.34	85.83	–
XPP	89.02	88.72	85.71
Supervised	88.80	85.43	–

Table 9: Recognition accuracy of same encoders (primary split, seed 1, %). Greyscale: test images converted to greyscale. 20-way: all test rock types at once, prototypes from other slides.

Figure 8: Left: share of PPL images whose nearest XPL image among all test fields is their own field. Right: cross-slide 1-shot accuracy on original (coloured dot) and greyscale (grey dot) test images.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

	Injected positives
Setting	correct	1-shot	5-shot
No replacement (own views only)	–	48.71	65.87
PETD	about 10%	49.47	64.47
PETD, false positives removed by oracle	100%	52.61	68.94
PETD, false negatives removed by oracle	about 10%	49.81	64.78

Appendix

Table A.1: Oracle decomposition of pseudo-task errors of PETD on miniImageNet (Conv4,

T=100

, 5-way test accuracy, %).

Training	Same-class pairs		False negatives
ways $N$	per batch	ProtoCLR	removed	Difference
5	0.16	40.89	41.17	$+0.28$
10	0.70	42.37	42.29	$-0.07$
20	2.96	42.98	43.20	$+0.22$
50	19.12	43.68	43.36	$-0.32$

Appendix

Table A.2: False negatives in ProtoCLR on miniImageNet (Conv4, 84

\times

84 images, 150 000 iterations, single seed, 5-way 1-shot test accuracy, %).

Precision of injected positives	Anchors using positives	1-shot	5-shot
0.30	100%	50.36	65.78
0.50	100%	51.63	67.23
0.70	100%	52.26	68.07
1.00	25%	50.94	67.78
1.00	50%	52.14	68.84

Appendix

Table A.3: Accuracy as a function of precision of injected positives and of share of anchors that receive them (miniImageNet,

T=100

, %). Precision is controlled with class labels.

Figure A.1: Test 5-way 1-shot accuracy of PETD on miniImageNet when injected positives are replaced by an oracle. Left: precision of injected positives at full coverage; right: share of anchors that receive correct positives.

Label-free selector	Precision of used neighbours	Query slots replaced (final epoch)	1-shot	5-shot
PETD (random co-clustered)	about 0.10	60%	49.47	64.47
Nearest co-clustered (kNN)	0.18	60%	49.76	65.83
Episodic prototype gate	0.22	17%	49.50	66.20
kNN + gate	0.25	32%	49.56	66.30
Global nearest neighbours	0.31	60%	49.74	66.59
Global NN, densest 50% of anchors	0.39	30%	49.17	66.09

Appendix

Table A.4: Label-free neighbour selectors on miniImageNet (

T=100

). Precision is measured on neighbours actually injected; replaced share is for final epoch.

Few-Shot Image Classification

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

See all 66 papers

FROST: Training-Free Few-Shot Segmentation with Frozen Features and Nonparametric Statistics
Jun 30, 2026Junghwan ParkRemote Sensing Image Segmentation Few-Shot Semantic Segmentation
TelePIX
Quantum-Enhanced Similarity Measures for Polarimetric Materials Classification
Jun 5, 2026Sara Shojaei, Seyed Mohamad Ali Tousi, Emma Bennett +6Hybrid Quantum-Classical ML Materials Science
University of Missouri–Columbia, MO, USA · Saint Louis University, MO, USA
Revisiting Shape from Polarization in the Era of Vision Foundation Models
Mar 5, 2026Chenhao Li, Taishi Ono, Takeshi Uemori +1Surface Normal Estimation Polarimetric Imaging
Sony Semiconductor Solutions Corporation

Jun 30, 2026J/K move · Enter open · S save

FROST: Training-Free Few-Shot Segmentation with Frozen Features and Nonparametric Statistics

Junghwan Park

TelePIX

Remote Sensing Image Segmentation Few-Shot Semantic Segmentation Semantic Segmentation

Open paper AskPDF arXiv

Few-shot segmentation asks a model to delineate a target class in a query image from only a handful of annotated examples, a setting most acute in remote sensing, where labels are scarce and the imagery departs sharply from the natural images on which vision backbones are pretrained. Prevailing approaches either train a segmenter on labelled episodes, which raises accuracy within the training distribution but binds the model to it, or reduce each class to a lossy summary of frozen features, a single prototype, a few cluster prototypes, or a discrete clustering, none of which preserves the internal structure of a multimodal class. We argue that a class is better described by a distribution than by a point, and that frozen self-supervised features already carry enough structure to estimate that distribution directly. We introduce FROST, a training-free few-shot segmenter that treats the reference foreground and background as two point clouds on the unit sphere of frozen DINOv3 features and labels each query token by a nonparametric density ratio, with a threshold the Bayes rule fixes at zero under equal priors. Because the variance of a density estimate shrinks as its sample grows, the decision sharpens as references accumulate, and every remaining quantity from the kernel bandwidth to the spatial gate is read from the support set rather than tuned. We develop FROST for overhead imagery, where a class is typically a scatter of many small and dissimilar instances that a density tracks but a lossy summary blurs. Across seventeen remote-sensing benchmarks FROST surpasses both training-free and learning-based methods, leading by 5.6 mIoU from a single annotated example and widening its lead as the support set grows, all while remaining among the smallest models compared. Code is available at https://github.com/jhpark-ai/FROST.

Abstract

Few-Shot Image Classification

Explore similar work

FROST: Training-Free Few-Shot Segmentation with Frozen Features and Nonparametric Statistics

Quantum-Enhanced Similarity Measures for Polarimetric Materials Classification

Revisiting Shape from Polarization in the Era of Vision Foundation Models