PaperlayerView exact revision on arXiv

Source-linked AI summary

Shortcut-Resistant CAM Distillation for Long-Tailed Recognition

Wenhai Wan, Teng Zhang, Shao-Yuan Li, Xinrui Wang, Qiang-Sheng Hua, Songcan Chen

arXiv:2610.29508v1cs.LG

TL;DR

Scarce and biased supervision makes tail-class generalization difficult and can make shortcut cues especially tempting. SRCD distills object-focused CAM explanations from head classes to tail classes, and the paper reports improved tail performance while maintaining head accuracy.

  • Problem

    Tail-class generalization remains difficult under scarce, biased supervision, which can make non-semantic cues appear reliable in context-homogeneous examples.

    Source: §1 · pp. 1–2

  • Method

    SRCD transfers object-focused explanations from head classes to tail classes through CAM distillation.

    Source: §1 · p. 2

  • Results

    SRCD improves tail-class performance while maintaining accuracy on head classes.

    Source: §6 · p. 9

  • Takeaways & Limitations

    The experiments support SRCD as a robust regularizer for long-tailed learning tasks.

    Source: §6 · p. 9

    SRCD’s training head–tail split uses a 100-example threshold, while split-wise evaluation reports separate Head, Med, and Tail groups.

    Source: §3.1 · p. 3

Abstract

from arXiv · show

Real-world datasets often follow a long-tailed distribution, making generalization to tail classes difficult. We revisit this problem through the lens of shortcut learning, where models prefer the easiest predictive cues (e.g., background or textures) over object-centric semantics, especially under scarce and biased supervision. We find that this tendency is amplified for tail classes: limited examples often share similar contexts, making non-semantic signals highly correlated and thus tempting shortcuts, whereas head classes with diverse appearances and environments encourage more stable object-focused representations. Motivated by this observation, we propose Shortcut-Resistant CAM Distillation (SRCD), a plug-and-play framework that transfers object-focused explanations from head to tail classes. SRCD operates in the Class Activation Map (CAM) space, where a CAM provides a class-specific spatial evidence map for a prediction. SRCD aggregates CAMs from a small set of head-class candidates into a shortcut-resistant teacher using an energy-model weighting based on coherence and concentration, and distills it to the tail-class CAM. We provide a theoretical analysis that quantifies shortcut reliance as shortcut-region evidence mass in CAM space and shows that SRCD suppresses tail shortcuts. Extensive experiments on long-tailed benchmarks consistently improve strong baselines. The code is available at https://github.com/Haifeng3/SRCD.

1. Introduction

Long-tailed recognition makes tail-class generalization difficult because scarce, context-homogeneous examples can make shortcuts appear reliable. SRCD transfers object-focused CAM explanations from head classes to tail classes, and the paper reports that it suppresses shortcut reliance and improves long-tailed benchmark performance.

1. Introduction

  • Scarce, context-homogeneous tail examples make non-semantic cues unusually reliable and can leave tail predictions shortcut-prone despite improved optimization or calibration.Head classes’ more diverse appearances and environments instead encourage stable object-focused evidence.Source: §1 · p. 2
  • SRCD transfers object-focused explanations from head-class CAMs to tail-class CAMs through a plug-and-play distillation strategy.It weights a small set of predicted head-class candidates by CAM coherence and concentration.Source: §1 · p. 2
  • The paper reports theoretical support for shortcut suppression and consistent performance improvements across long-tailed benchmarks.SRCD is presented as complementary to existing long-tailed learning methods.Source: §1 · p. 2

2. Related Work

Prior long-tailed methods address scarce tail data through sample, feature, or distributional-statistics transfer, while shortcut learning has been studied mainly in standard recognition settings. The paper focuses on shortcut learning under long-tailed distributions and intervenes at the evidence level.

2. Related Work

  • Long-tailed methods address tail scarcity by synthesizing samples or features, or transferring distributional statistics and contexts.Examples include M2m, RSG, FASA, LTA, CMO, and H2T.Source: §2 · p. 2
  • Shortcut learning describes reliance on predictive but spurious cues such as background context or textures instead of semantic evidence.Prior work in standard recognition has studied these shortcuts and proposed augmentation and bias-mitigation interventions.Source: §2 · p. 2
  • The role of shortcut learning under long-tailed distributions remains largely unexplored, although scarce and biased tail supervision can make shortcuts unusually consistent and easy to fit.The paper distinguishes its approach by intervening on spatial evidence rather than samples, features, or statistics.Source: §2 · pp. 2–3

3. Methodology

The methodology defines the long-tailed learning and evaluation setup, uses CAMs to inspect class-specific spatial evidence, and diagnoses a head–tail gap against balanced-model explanations. SRCD then selects and weights head-class proxy CAMs to form a teacher that is distilled into tail-class CAMs.

3.1. Problem Setup

  • The setup trains on an imbalanced distribution and targets low risk on a balanced test distribution, with head and tail groups split at 100 training examples.The balanced and long-tailed distributions are assumed to share class-conditional distributions; split-wise evaluation instead reports Head, Med, and Tail groups.Source: §3.1 · p. 3

3.2. Revisiting Class Activation Maps

  • CAMs expose class-specific spatial evidence, and normalized CAMs represent the fraction of class evidence allocated to each image location.This lets the method compare attention patterns and assess whether evidence falls on meaningful objects or shortcut cues.Source: §3.2 · p. 3

3.3. Diagnosing Shortcut Learning in Long-Tailed Recognition

  • Long-tailed-model CAM similarity to a balanced-model reference drops on tail classes, while head-proxy CAMs are more similar than tail-class CAMs for tail examples.Figure 2 compares ground-truth-class CAMs and head-proxy CAMs against balanced-model CAMs; the reference is treated as a proxy for semantic explanations.Source: §3.3 · pp. 3–4, Figure 2 · p. 4

3.4. Shortcut-Resistant CAM Distillation

  • For each tail example, SRCD selects the top-K predicted head classes and treats their CAMs as proxy explanations that can provide more stable spatial evidence.The method does not require strict semantic matching or explicit part-level decomposition.Source: §3.4 · p. 4
  • SRCD scores proxy CAMs for spatial coherence and evidence concentration, then energy-weights them to favor smooth, compact explanations over fragmented or diffuse ones.Coherence uses a Laplacian energy, while concentration uses a normalized sparsity measure.Source: §3.4 · p. 5
  • An L2 distillation loss aligns each tail-class CAM with the teacher, using stronger class-dependent regularization for rarer classes alongside a long-tailed classification loss.SRCD is architecture-agnostic, can accompany existing methods, and is enabled in late training when head explanations are more reliable.Source: §3.4 · pp. 5–6

4. Theoretical Analysis

The analysis defines shortcut reliance as CAM mass on shortcut regions and explains how energy-weighted teacher construction and L2 distillation reduce that mass in tail CAMs.

4. Theoretical Analysis

  • SRCD defines shortcut reliance as normalized CAM evidence mass assigned to a shortcut/background region.Source: §4 · p. 6
  • The teacher’s utility weighting favors head-proxy CAMs with low coherence energy and high concentration while entropy regularization prevents brittle selection.The softmax weights are characterized as the unique maximizer of a strictly concave objective.Source: §4 · p. 6
  • The distillation weight γ_y increases with tail scarcity, strengthening teacher-to-tail CAM distillation for scarcer classes.Source: §4 · p. 6
  • Under L2 distillation, when the initial tail CAM has more shortcut mass than the teacher, that mass decreases monotonically and converges exponentially to the teacher’s mass.The analysis describes the CAM’s shortcut attention as contracting toward the teacher at the same rate as the full CAM.Source: §4 · p. 6

5. Experiments

SRCD is evaluated as a plug-and-play addition to long-tailed learning methods, both from scratch and when fine-tuning foundation models. It consistently improves results across the reported settings, with ablations examining distillation strength, proxy count, and teacher weighting.

5.1. Experiment Protocols

  • Experiments cover five long-tailed benchmarks and compare SRCD with representative methods under scratch training and foundation-model fine-tuning.The scratch setup uses ResNet backbones, while fine-tuning follows a CLIP adaptation protocol.Source: §5.1 · pp. 6–7

5.2. Results with training from scratch

  • From-scratch SRCD improves overall accuracy across evaluated methods and datasets, with particularly strong gains on Medium/Tail splits for ImageNet-LT and iNaturalist.The reported comparisons include CIFAR-10-LT and CIFAR-100-LT across imbalance ratios, plus large-scale benchmarks.Source: §5.2 · p. 7, Table 1 · p. 7, Table 2 · p. 8

5.3. Results with fine-tuning foundation models

  • Fine-tuning SRCD consistently improves overall accuracy on ImageNet-LT and iNaturalist, with Medium/Tail gains alongside stable or improved Head performance.It is added to strong fine-tuning baselines without modifying architectures.Source: §5.3 · p. 8, Table 2 · p. 8

5.4. Ablation Study

  • When β exceeds 0.6, accuracy improves significantly and stably; K = 3 provides most gains, with larger K yielding marginal improvements.β controls distillation strength, and K is the number of head-class proxies used to build the teacher.Source: §5.4 · pp. 8–9
  • Using both coherence and concentration cues gives the best performance, while energy-based aggregation yields higher CAM similarity to a balanced reference than simple averaging.The weighting comparison indicates that the resulting teacher is more shortcut-resistant.Source: Table 3 · p. 8, §5.4 · p. 9
  • Replacing the head-proxy teacher with a fixed center mask degrades performance, especially on Tail, whereas SRCD improves the baseline.Computational overhead and qualitative CAM visualizations are reported in the Appendix.Source: §5.4 · p. 9

6. Conclusion

SRCD transfers reliable spatial evidence from head classes to tail classes to reduce shortcut learning, improving tail-class performance while maintaining head-class accuracy.

6. Conclusion

  • SRCD improves tail-class performance while maintaining head-class accuracy by distilling reliable spatial evidence from head classes.The conclusion presents SRCD as a robust regularizer for long-tailed learning tasks.Source: §6 · p. 9

A.1. Long-tailed Recognition and Head-to-Tail Transfer

  • Head-to-tail transfer addresses tail scarcity through semantic knowledge transfer, synthesized samples or features, and feature fusion.These approaches aim to enrich tail representations with diversity missing from scarce examples.Source: §A.1 · p. 12

A.2. Shortcut Learning and Spurious Correlations

  • With limited supervision, models can favor easy predictive cues such as background or texture over object-centric semantics, weakening out-of-distribution generalization.Context and texture biases are common visual manifestations of shortcut learning.Source: §A.2 · pp. 12–13
  • Long-tailed supervision can make tail contexts unusually predictive, while diverse head-class examples weaken spurious cues and encourage stable object evidence.The paper connects this head–tail difference to shortcut-dominated CAMs in tail classes.Source: §A.2 · p. 13, §B.1 · p. 13

B.1. Additional Discussion on CAM-Based Shortcut Diagnosis

  • The diagnostic treats a balanced model as a practical semantic reference and interprets low CAM similarity as consistent with shortcut-dominated attention, not definitive proof.CAM normalization reduces scale effects, while cosine similarity compares spatial evidence patterns.Source: §B.1 · pp. 13–14
  • SRCD itself does not require a balanced reference model at deployment; head-class CAMs instead provide a practical source of shortcut-resistant evidence for tail classes.The balanced reference is used only to diagnose the shortcut phenomenon.Source: §B.1 · p. 14

B.2. Additional Discussion on the Rationale of Eh and En

The discussion motivates spatial coherence and evidence concentration as complementary CAM criteria for reducing shortcut reliance, while noting cases where these measures may penalize valid evidence.

Why spatial coherence is important

  • The Laplacian-based coherence energy Eh penalizes fragmented activations and favors smoother, more contiguous CAM evidence.The paper associates fragmented patterns with incidental background cues and shortcut-driven explanations.Source: §B.2 · p. 14

The role of evidence concentration

  • The concentration score En measures whether evidence is focused in a small region, discouraging diffuse attention that may cover background.The discussion links higher concentration with more object-centric explanations.Source: §B.2 · p. 14

Why these metrics work together

  • Eh and En address complementary patterns: coherence penalizes compact but fragmented peaks, while concentration penalizes smooth evidence spread across a large area.Combining them is intended to produce a teacher CAM that is both coherent and concentrated.Source: §B.2 · p. 15

Limitations of the approach

  • The metrics may penalize valid evidence for large or irregular objects, naturally disconnected parts, and fine-grained categories.The discussion notes these measures are not perfect proxies for object-centric semantics.Source: §B.2 · p. 15

Practical considerations

  • The method computes Eh and En on normalized CAMs and reuses a resolution-specific Laplacian; its 4-neighbor grid could be replaced for more complex tasks.The discussion reports that the 4-neighbor grid works well on most datasets.Source: §B.2 · p. 15

Conclusion

  • SRCD combines spatial coherence and evidence concentration to suppress shortcut-driven behavior in tail classes, with particular relevance when background cues dominate.The paper describes the metrics as a regularizer for tail-class predictions and reports generalization across long-tailed benchmarks.Source: §B.2 · p. 15

B.3. Discussion on Head-Proxy CAM Supervision

SRCD treats head-proxy CAMs as a regularizing teacher for tail explanations, not as exact semantic or part-level labels. It weights multiple proxies by spatial coherence and evidence concentration, and applies the distillation specifically to tail classes.

B.3. Discussion on Head-Proxy CAM Supervision

  • SRCD uses head-proxy CAMs as a regularizer to move tail explanations away from context shortcuts, without requiring perfect object-part coverage or alignment.The teacher need only be more stable and less shortcut-prone than the original tail CAM; proxies need not cover complementary parts.Source: §B.3 · p. 16
  • Weighted aggregation favors proxy CAMs with coherent, concentrated spatial evidence, reducing dependence on any single potentially mismatched candidate.The paper identifies coherence with Laplacian smoothness and concentration with Hoyer sparsity.Source: §B.3 · p. 16, §C.1 · p. 17
  • On taxonomically mismatched iNaturalist tail samples, SRCD raises BCL accuracy from 67.25 to 69.88; the subset comprises 32.13% of tail samples.The analysis indicates that exact semantic matching between proxy and target is not required.Source: §B.3 · p. 16, Table 4 · p. 16

B.4. Long-Tailed-Specific Design of SRCD

  • SRCD is tailored to long-tailed imbalance: head proxies regularize tail attention, with correction strength scaled by class rarity.The method is presented as distinct from generic CAM distillation applied to long-tailed data.Source: §B.4 · p. 17

C.1. Notation and Preliminaries

  • Normalized CAMs are treated as spatial probability distributions, and shortcut reliance is quantified as evidence mass assigned to a conceptual shortcut/background region.The region partition formalizes the analysis and does not require explicit masks during training.Source: §C.1 · p. 17

C.2. Proof of Theorem 4.1 (Entropy-Regularized Utility Maximization)

  • Entropy-regularized utility weighting selects a unique distribution over head-proxy candidates, balancing higher-utility candidates against brittle hard selection.Utility rewards coherent, concentrated evidence; temperature controls the trade-off between near-hard selection and uniform averaging.Source: §C.2 · p. 18

C.3. Proof of Theorem 4.2 (Distillation Dynamics)

The distillation dynamics interpolate the tail CAM toward the teacher, transporting shortcut-region mass toward the teacher’s level. The teacher’s expected shortcut resistance is supported under a stated negative alignment between utility and shortcut mass.

C.3. Proof of Theorem 4.2 (Distillation Dynamics)

  • If the teacher has less shortcut mass than the initial tail CAM, gradient-flow distillation reduces shortcut mass monotonically toward the teacher’s level.The CAM follows an exponential interpolation toward the teacher.Source: §C.3 · p. 19
  • A gradient step interpolates the tail CAM toward the teacher, so shortcut mass decreases when the teacher has lower shortcut mass.The same transport follows from linearity of shortcut-region mass.Source: §C.3 · p. 19

C.4.1. A WEAK ALIGNMENT ASSUMPTION

  • The teacher analysis assumes utility is negatively aligned with shortcut mass: candidates with higher utility have lower shortcut mass under the stated condition.The proposed utility favors coherent and concentrated evidence.Source: §C.4.1 · p. 20

C.4.2. TEACHER SHORTCUT MASS BOUND UNDER ALIGNMENT

  • Under that alignment, teacher shortcut mass is bounded by negative expected utility, so increasing expected utility reduces or controls the bound.Utility combines a penalty for fragmentation with a reward for concentration.Source: §C.4.2 · p. 20

C.4.3. SOFTMAX WEIGHTS MAXIMIZE EXPECTED UTILITY (WITH ROBUSTNESS)

The appendix connects utility-weighted teacher construction to shortcut-mass reduction, while detailing the coherence and concentration criteria used to select CAMs. The analysis also identifies training timing and assumptions that delimit the mechanism.

C.4.3. SOFTMAX WEIGHTS MAXIMIZE EXPECTED UTILITY (WITH ROBUSTNESS)

  • Entropy-regularized softmax weights favor higher-utility candidates while avoiding collapse to one candidate when utilities are close or noisy.The weights maximize expected utility with an entropy-based robustness term.Source: §C.4.3 · p. 20

C.4.4. FROM TEACHER QUALITY TO TAIL SHORTCUT SUPPRESSION

  • If the teacher has lower shortcut mass than the tail CAM, distillation monotonically reduces tail shortcut mass toward the teacher.The appendix describes the gradient-flow reduction as exponential.Source: §C.4.4 · p. 20

C.5. Technical Details of Eh: Graph Laplacian Coherence Energy

  • Graph-Laplacian coherence energy penalizes neighboring-pixel discontinuities, which are associated with fragmented shortcut activations rather than connected object evidence.The spectral account says minimizing this energy suppresses high-frequency CAM oscillations.Source: §C.5.1 · p. 21, §C.5.2 · p. 21

C.5.3. DISCUSSION: COHERENCE ENERGY VS. “PURE CONNECTIVITY”

  • Compared with thresholded connectivity counts, Laplacian coherence energy is differentiable, stable to small perturbations, scale-aware, and locally computable.The paper presents it as a practical surrogate for shortcut-related fragmentation.Source: §C.5.3 · p. 21

C.6. Technical Details of En: Evidence Concentration Score

  • Hoyer-style concentration rewards CAM evidence concentrated in fewer locations, countering diffuse background spread; the appendix notes that coherence and concentration scales may need matching.The suggested weighted utility is U = −λhEh + λnEn, with positive coefficients.Source: §C.6.1 · p. 22, §C.6.2 · p. 22

C.7. Additional Bounds: Shortcut Mass Gap Controlled by Distillation Distance

  • A static bound links shortcut-mass discrepancy to CAM distillation distance, complementing the dynamic transport analysis.The bound is presented as a worst-case guarantee without analyzing learning dynamics.Source: §C.7 · p. 22

C.8. Why SRCD Should be Enabled After Head Classes Become Reliable

  • SRCD should start after head representations become reliable, because early noisy proxy utilities may fail to reduce shortcut mass and could amplify shortcuts.The paper aligns this late start with second-stage fine-tuning protocols.Source: §C.8 · p. 22

C.9. Extended Discussion and Practical Takeaways

  • The theory treats shortcuts as CAM mass on a conceptual region and requires neither segmentation masks nor explicit shortcut-region knowledge during training.The appendix summarizes SRCD as evidence-level suppression, not part-level decomposition, with rarity-weighted correction for tail classes.Source: §C.9 · p. 23

D.1. Detailed Dataset Description

The experiments use long-tailed benchmarks spanning synthetic imbalance variants and real-world datasets. SRCD is evaluated as an architecture-preserving add-on against both training-from-scratch methods and foundation-model fine-tuning approaches.

D.1. Detailed Dataset Description

  • The benchmark suite includes CIFAR100-LT variants with different imbalance ratios, ImageNet-LT, iNaturalist 2018, and Places-LT.The text characterizes the datasets as severely imbalanced; iNaturalist has 8,142 categories and 437.5K training images.Source: §D.1 · p. 24

D.2. Competitors

  • SRCD is added without changing baseline architectures or training hyperparameters and is enabled after warm-up in the late training phase.The reported experiments use the corresponding baseline training pipeline.Source: §D.2 · p. 24

D.2.1. TRAINING FROM SCRATCH

  • The training-from-scratch comparisons include optimization, contrastive-learning, uncertainty, feature-geometry, consistency, transfer, and two-stage methods.Examples listed include Focal-SAM, BCL, GCL, FCC, GLMC, H2T, and BNS.Source: §D.2.1 · pp. 24–25

D.2.2. FINE-TUNING FOUNDATION MODELS

  • For vision-language adaptation, SRCD is compared with decoder, expert-learning, and lightweight foundation-model adaptation approaches.The paper describes SRCD as orthogonal and plug-and-play across these baseline settings.Source: §D.2.2 · pp. 25–26

D.3. Additional Results

Additional benchmark results show SRCD improving tail accuracy across long-tailed settings, while implementation details explain how CAM supervision is constructed for different encoders. The section also compares adaptive teacher construction with a fixed center prior and reports a small training-time increase.

D.3. Additional Results

  • Under CIFAR100-LT IR=100, SRCD improves Head, Medium, Tail, and overall accuracy for every evaluated baseline; under IR=50, it improves Tail accuracy for all baselines, despite a slight overall drop for CE+H2T.The most pronounced IR=100 gains are on the Tail split.Source: Table 6 · p. 26, Table 7 · p. 26, §D.3 · p. 26
  • On Places-LT, SRCD improves overall accuracy for all evaluated scratch baselines and improves Tail accuracy in both scratch and foundation-model fine-tuning settings.Tail accuracy gains for BCL, GLMC, GCL, and H2T are 1.32, 1.84, 0.58, and 2.96 percentage points; fine-tuning Tail gains for PECEL, Decoder, and LIFT are 0.9, 1.5, and 2.3 points.Source: §D.3 · pp. 26–27, Table 8 · p. 27

D.4. CAM Extraction Details for Fine-tuning Foundation Models

  • For fine-tuning, CAMs use classifier weights and spatial features, with ViTs represented as patch-token grids and CNNs using final convolutional feature maps.The target CAM is conditioned on the ground-truth class.Source: §D.4.1 · p. 27, §D.4.2 · p. 28, §D.4.3 · p. 28

D.4.4. CLIP-SPECIFIC CONSIDERATIONS (VISION-LANGUAGE ENCODERS)

  • In CLIP fine-tuning, a learnable classification head supplies CAM weights, aligning the spatial evidence with the classifier used at evaluation.For ViTs, the patch-token features exclude the class token and are reshaped into a spatial grid.Source: §D.4.4 · p. 28, §D.4.2 · p. 28

D.4.5. STOP-GRADIENT ON THE CAM BRANCH

  • The CAM branch uses detached spatial features during fine-tuning, preventing CAM supervision from directly reshaping backbone features while retaining an auxiliary SRCD signal.The signal can act through designated parameters such as the SRCD projection or adapters.Source: §D.4.5 · p. 29

D.4.6. MULTI-CAM COLLECTION FOR SRCD TEACHER CONSTRUCTION (FINE-TUNING)

  • SRCD builds a teacher by weighting a fixed set of head-proxy CAMs; fine-tuning experiments use K = 3 proxies and loss weight β = 0.6.The teacher is described as typically more object-centric and less sensitive to long-tailed noise.Source: §D.4.6 · p. 29, §D.4.8 · p. 29

D.5. Comparison with a Center-Mask Prior

  • On CIFAR100-LT, the fixed center-mask baseline performs worse than BNS, especially on Tail; SRCD instead uses an adaptive head-proxy teacher weighted for coherent, concentrated evidence.The center-mask comparison is under IR=100.Source: §D.5 · pp. 29–30, Table 9 · p. 30

D.6. Computational Overhead

  • Adding SRCD causes only a small increase in iteration time, while total training time remains close to the baseline; late-stage activation and reuse of forward features help limit overhead.CAM construction also avoids attribution backpropagation and uses a small fixed proxy count.Source: §D.6 · p. 30, Table 10 · p. 30

D.7. Visualization of CAM Evidence

The visualizations compare head-proxy CAMs with tail ground-truth CAMs and show how SRCD changes attention patterns. Proxy candidates can contribute complementary object evidence, while SRCD produces more object-focused and spatially coherent CAMs.

Head-proxy CAMs vs. tail ground-truth CAM

  • Among head-proxy CAM candidates, some align with tail ground-truth object evidence while others highlight scattered or background regions; aggregation favors coherent, concentrated candidates.Different candidates can emphasize complementary object parts, while their spurious responses are less consistent.Source: Figure 6 · p. 31, §D.7 · p. 31

Effect of SRCD on CAM quality

  • With SRCD, CAMs become more concentrated on objects, cover multiple object-relevant regions, and show more consistent attention across examples than without SRCD.The comparison is shown across multiple examples.Source: §D.7 · p. 31, Figure 7 · p. 32

D.8. Direct Evidence for Shortcut Suppression on VOC

On a curated long-tailed VOC 2012 subset, SRCD increases CAM evidence within foreground masks relative to background and improves accuracy. The section also describes the training procedure that constructs and distills a weighted head-proxy teacher for tail samples.

D.8. Direct Evidence for Shortcut Suppression on VOC

  • The VOC analysis uses LIFT as baseline on a curated single-label subset with imbalance ratio 80; FG/BG Ratio measures CAM evidence inside versus outside the segmentation mask.A larger ratio indicates more object-region evidence and less reliance on background shortcuts.Source: §D.8 · p. 33
  • FG/BG Ratio rises from 0.85 to 1.56, and accuracy from 78.37 to 83.16, on long-tailed VOC 2012 with SRCD.The higher ratio indicates more CAM evidence inside the foreground mask relative to background.Source: §D.8 · p. 33
  • Figure 8 shows SRCD CAMs more concentrated on foreground objects and less dominated by background context than baseline CAMs.This qualitative comparison is consistent with the mask-based analysis.Source: §D.8 · p. 33, Figure 8 · p. 34

E. Algorithm

  • For each tail sample, the algorithm computes a student tail CAM and detached head-proxy CAMs, weights proxies, forms a detached teacher, and optimizes classification plus SRCD loss.The algorithm selects top-K predicted head classes for proxy CAM construction.Source: E. Algorithm · p. 35

Related papers

Loading 2610.29508v1…