cs.CVSep 24, 2026

The Shape of Events: Edge-Based Inductive Biases via Cross-Domain Distillation

Authors: Soshun Kihara, Shunsuke Yasuki, Masato Taki

Organizations: Independent Researcher · The University of Tokyo, Japan · Graduate School of Artificial Intelligence and Science, Rikkyo University, Japan · RIKEN, Japan

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Convolutional neural networks trained on ImageNet are known to exhibit a strong preference for local high-frequency texture, an inductive bias that translates into fragile robustness against distribution shifts in real-world environments. Event cameras, in contrast, record only changes in scene brightness and are therefore well suited to capturing contour information; however, due to the absence of diagnostic benchmarks in the event domain, the inductive bias that event-camera data instills in vision models has remained underexplored. In this work, we use knowledge distillation from the event domain to the RGB domain so as to exploit the rich evaluation toolkit available in the RGB domain and systematically dissect this inductive bias. Our experiments show that distillation from the event domain induces, in the RGB domain, color invariance, shape bias, and robustness to high-frequency noise. We identify the underlying mechanism as the model suppressing its dependence on high-frequency texture while acquiring a stronger dependence on edge-based object shape. This hypothesis is supported by changes in how color and spatial information are processed at the early layers, together with a spectral trade-off in which robustness to the absence of high-frequency components coexists with vulnerability to contamination of the relied-upon frequency bands and to disruption of geometric structure. We further show that this inductive bias differs from existing robustification methods and that it functions as a useful prior for diverse downstream tasks in which shape and contour information contribute alongside other cues. The code is available at https://github.com/snskysk/event2rgb-distillation .

Figures & tables

Figure 1: Method overview and summary of results. Left : Cross-domain knowledge distillation from a frozen DiST event teacher (trained on N-ImageNet) into an RGB ResNet-34 student (trained on ImageNet). N-ImageNet and ImageNet are paired at the instance level, enabling per-sample KL between teacher and student logits. Right : Summary of robustness gains of ED_Student ( α=0.2\alpha=0.2 ) over the RGB baseline across the three axes studied in this paper (color invariance, shape bias, and high-frequency noise robustness) and downstream task transfer.
Figure 2: Robustness properties emerging from event-driven knowledge distillation. a) Top-1 accuracy across hue rotation angles. b) Cue-conflict shape bias; Event-distilled students approach the human reference. c) Accuracy under PGD attacks with 0<ϵ≤1/2550<\epsilon\leq 1/255 .
Figure 3: Internal-representation analysis and the spectral trade-off. a) First-layer filter excerpts, each min-max normalized to [0,1][0,1] : baseline shows noisy high-frequency patterns; the Event-distilled model (especially ED_SoftLabelOnly) acquires smooth monochromatic, color-opponent, and Gabor-like structures. b) RGB and spatial autocorrelations; ED_SoftLabelOnly exceeds the baseline on both, while ED_Student stays close to it. c) Two ImageNet-C corruptions, as accuracy retention (noisy / clean). Left : under JPEG compression, the Event-distilled model retains more at high severity. Right : under pixelate, a reversal occurs at high severity.
Figure 4: Ablation (Section 6.1 ). Adversarial robustness of color-stripped controls (Self / Gray / B&W Distilled) vs. the Event-distilled models. Left : FGSM. Right : PGD. x-axis: perturbation magnitude ϵ\epsilon ; y-axis: top-1 accuracy. The Event-distilled model surpasses all controls at 0<ϵ≤1/2550<\epsilon\leq 1/255 , indicating its inductive bias is not merely the result of color removal.
Figure 5: Architectural generality (Section 6.2 ). Comparison of baseline / ED_Student / ED_SoftLabelOnly across ResNet-34, VGG-16, ConvNeXt-Base, ResNet-50, and RepLKNet31B [ 79 ] . Top: FGSM. Bottom: PGD. x-axis: perturbation magnitude ϵ\epsilon . ED_Student improves over the baseline in 19 of the 30 tested (architecture, attack, ϵ\epsilon ) settings with 0<ϵ≤1/2550<\epsilon\leq 1/255 . The 11 exceptions, reported as the ED_Student −- baseline difference in percentage points, are (RepLKNet31B, FGSM) at ϵ=0.25,0.5,1\epsilon=0.25,0.5,1 /255: −3.9-3.9 , −11.9-11.9 , −20.7-20.7 ; (RepLKNet31B, PGD) at ϵ=0.5,1\epsilon=0.5,1 /255: −4.5-4.5 , −9.4-9.4 ; (ConvNeXt-Base, PGD) at ϵ=0.25,0.5,1\epsilon=0.25,0.5,1 /255: −15.6-15.6 , −10.1-10.1 , −2.8-2.8 ; (ConvNeXt-Base, FGSM) at ϵ=0.25,1\epsilon=0.25,1 /255: −0.4-0.4 , −0.6-0.6 ; and (ResNet-50, FGSM) at ϵ=1\epsilon=1 /255: −1.2-1.2 .
 C100C10ChXRMNISTCarsDTDSTL10INet-skCal101INet-v2SceneSUN397FlwrKITTIESATCUBPCamPetFoodMeanINet
base. (%)55.878.087.389.946.960.194.862.285.449.690.653.585.373.490.064.678.090.561.473.573.9
ED ( Δ\Delta )5.24.84.33.42.31.91.81.41.20.90.80.80.70.70.40.30.2−- 0.1−- 0.41.6−- 4.2
Adv. ( Δ\Delta )14.610.82.95.6−- 8.7−- 9.1−- 0.7−- 5.20.4−- 12.3−- 0.5−- 7.8−- 9.42.1−- 1.2−- 3.40.0−- 3.2−- 9.5−- 1.8−- 7.9
Noise ( Δ\Delta )3.22.60.6−- 0.7−- 5.6−- 2.50.6−- 4.50.71.7−- 0.8−- 2.3−- 3.90.4−- 1.70.20.3−- 0.1−- 3.8−- 0.80.5
SIN ( Δ\Delta )6.25.60.01.8−- 8.3−- 4.21.4−- 4.50.7−- 4.1−- 0.3−- 1.9−- 5.53.2−- 2.2−- 3.0−- 1.2−- 0.6−- 1.1−- 0.9−- 1.7
Canny ( Δ\Delta )−- 1.81.00.20.31.11.40.91.5−- 0.12.90.20.20.1−- 0.1−- 1.7−- 2.10.30.1−- 0.70.2−- 2.8
Table 1: Linear probe transfer (Sections 6.3 , 6.4 ). base. gives baseline top-1 (%); the Δ\Delta rows (%, ++ omitted) are ED_Student, Adversarial Training, Noise Training, SIN+IN co-training [ 1 ] , Canny_Distilled, and Synth_Distilled. Mean averages the 19 transfer tasks; INet is the pre-training reference, excluded from Mean. Bold: column-wise best; datasets sorted by ED_Student’s Δ\Delta ; abbreviations expanded in Appendix A.5 .
 CleanΔ\Delta Color (hue Avg.)Δ\Delta Shape (sketch)Δ\Delta High-freq (bandpass)Δ\Delta Transfer (LP Avg.)Axes improved
baseline (abs.)73.961.823.729.673.5—
ED_Student69.7+3.6+3.6+4.4+4.4+17.4+17.4+1.6+1.64 / 4
Canny_Distilled71.1+4.3+4.3+4.2+4.2−5.9-5.9+0.2+0.23 / 4
Synth_Distilled71.8−3.4-3.4−3.2-3.2+18.9+18.9−5.4-5.41 / 4
Adv. Training66.0−11.2-11.2−2.4-2.4+28.1+28.1−1.8-1.81 / 4
Noise Training74.4+0.8+0.8+1.7+1.7+13.4+13.4−0.8-0.83 / 4
Table 2: Event supervision vs. five non-event supervisions, all with the same ResNet-34 student. Clean is absolute for every row; on the four axes the baseline row is absolute and the others are differences from it, signed so that positive means better. Color: mean top-1 over ten hue-rotation angles; Shape: ImageNet-Sketch top-1 [ 85 ] ; High-freq: reduction in the accuracy drop under band-limited noise at 0.95π0.95\pi , where the baseline drops 29.6 pp; Transfer: mean linear-probe accuracy over the 19 tasks of Table 1 . Only ED_Student improves on all four axes.
Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

ItemValue
Input representationDiST ( reshape_then_acc_adj_sort ), 2 ch, 224×224224\times 224
Event slicingRandom, slice_length =30,000=30{,}000
Batch size512
Weight decay1×10−41\times 10^{-4}
Momentum0.9
Epochs100
Appendix
Table 3: Shared training recipe for the in-house DiST teachers of Section 6.2 . Architecture-specific settings are in Table 4 . The DiST ResNet-34 teacher used elsewhere is the publicly released checkpoint of [ 5 ] and is not retrained by us.
ArchitectureOptimizerLearning rateFirst-conv kernel size
ResNet-50Adam2×10−42\times 10^{-4}14
VGG16-BNAdam2×10−42\times 10^{-4}14
ConvNeXt-BaseAdamW2×10−42\times 10^{-4}4
RepLKNet31BAdamW2×10−42\times 10^{-4}4
Appendix
Table 4: Architecture-specific settings for the DiST teachers in Section 6.2 . Common settings are in Table 3 . The DiST ResNet-34 teacher uses the public checkpoint of [ 5 ] and is omitted.
Modelα\alphaTeacher
baseline0.0–
ED_Student (main)0.2DiST ResNet-34
ED_SoftLabelOnly1.0DiST ResNet-34
Self Distilled0.2RGB ResNet-34 baseline
Gray Distilled0.2Grayscale ResNet-34
B&W Distilled0.2B&W ResNet-34
Appendix
Table 5: Distillation weight α\alpha (Eq. 1 ) and the teacher used for each ResNet-34 student. The KL-term temperature is fixed at T=3.0T=3.0 . Adversarial / Noise Training and the SIN model use only a CE loss and are omitted from this table (see Section A.3.3 , A.3.4 , respectively).
ModelLRWeight decayEpochsLR schedule
baseline / ED_Student / ED_SoftLabelOnly2×10−32\times 10^{-3}5×10−25\times 10^{-2}1205-epoch linear warmup →\rightarrow cosine
Self / Gray / B&W Distilled7.5×10−47.5\times 10^{-4}5×10−25\times 10^{-2}1205-epoch linear warmup →\rightarrow cosine
SIN2×10−32\times 10^{-3}5×10−25\times 10^{-2}1205-epoch linear warmup →\rightarrow cosine
Adversarial Training1×10−31\times 10^{-3}2×10−42\times 10^{-4}50Cosine annealing (no warmup)
Noise Training1×10−31\times 10^{-3}2×10−42\times 10^{-4}120Cosine annealing (no warmup)
Appendix
Table 6: Per-student optimization configuration for the ResNet-34 student models.
OperationSetting
RandomResizedCropoutput 224224 , scale =(0.08,1.0)=(0.08,1.0)
RandomHorizontalFlipp=0.5p=0.5
RandAugmentnum_ops =2=2 , magnitude =9=9
NormalizeImageNet mean/std
RandomErasingp=0.25p=0.25
Appendix
Table 7: RGB strong augmentation pipeline used on the CE branch of every ResNet-34 student. Adversarial Training omits RandAugment [ 93 ] and RandomErasing [ 94 ] to avoid interference with PGD noise.
OperationSetting
Resize256
CenterCrop224
NormalizeImageNet mean/std
Appendix
Table 8: RGB weak augmentation pipeline used on the KD branch (student input during KD loss computation) and on every validation loader.
OperationSetting
RandomResizedCropoutput 224224 , scale =(0.08,1.0)=(0.08,1.0) , NEAREST interpolation
RandomHorizontalFlipp=0.5p=0.5
NormalizeImageNet mean/std
Appendix
Table 9: Augmentation pipeline for binarized (B&W) inputs. All spatial operations use NEAREST interpolation to preserve the two-valued pixel distribution; RandAugment and RandomErasing are omitted.
 TeacherStudent (ImageNet)
ArchitectureDiSTED_SoftLabelOnlyED_Studentbaseline
ResNet-3448.4353.5069.7073.90
VGG16-BN42.1324.1455.9069.01
ConvNeXt-Base52.4270.9778.4081.50
ResNet-5051.1159.6867.4778.16
RepLKNet31B51.6167.4376.6183.45
Appendix
Table 10: Clean top-1 accuracy (%) of the teacher and student models used in the main comparisons. The Teacher column (DiST) is evaluated on N-ImageNet validation (event domain); the Student columns are on ImageNet validation (RGB domain). The two additional non-event controls are listed separately in Table 11 .
ControlTeacher inputTeacherStudent
Canny_DistilledCanny edge map (3ch)55.8271.15
Synth_Distilledaugmentation-difference ON/OFF (2ch)64.5271.78
Appendix
Table 11: Clean top-1 accuracy (%) of the two additional non-event controls. Teacher accuracy is measured on the correspondingly transformed ImageNet validation split; student accuracy is on the standard RGB validation split.
 SOTA baselineSOTA ED_Student
Clean top-175.9270.80
ImageNet-Sketch top-126.9828.94
Hue rotation, mean over 3636 – 324∘324^{\circ}63.9665.76
Patch shuffle, grid =4=453.8237.48
Patch rotation, grid =4=446.5431.61
FGSM, mean over ϵ∈{0.25,0.5,1}/255\epsilon\in\{0.25,0.5,1\}/25534.7428.70
Appendix
Table 12: Baseline and ED_Student retrained under the SOTA recipe. ED’s advantage persists on shape, color, high-frequency-band robustness and downstream transfer, but not on the adversarial axes, where the stronger augmentation already makes the SOTA baseline robust at these perturbation sizes.
 LP meanHue Avg.GrayscaleCue-conflictSketchBandpass @ 0.95π0.95\piAdv. AUC
baseline73.561.862.40.21223.744.36.2
ED_Student75.165.464.60.37528.157.58.9
ED_SoftLabelOnly—52.452.60.48724.545.89.1
Canny_Distilled73.766.165.10.27027.935.67.5
Adversarial Training71.750.649.50.57221.364.524.1
Noise Training72.762.663.30.25825.458.211.5
Appendix
Table 13: Comparison on the axes of Figure 1 . LP mean: linear-probe accuracy averaged over the 19 tasks of Table 1 ; the linear probe is run for the five models compared in that table plus Canny_Distilled. Hue Avg.: top-1 averaged over ten hue-rotation angles including 0∘0^{\circ} . Grayscale: top-1 on the grayscale validation split. Cue-conflict: shape bias under the 16-class restricted-argmax protocol. Sketch: ImageNet-Sketch top-1. Bandpass: top-1 under band-limited noise centered at 0.95π0.95\pi (higher is better). Adv. AUC: area under the PGD accuracy curve over ϵ∈{0,0.25,0.5,1}/255\epsilon\in\{0,0.25,0.5,1\}/255 , scaled by 100. Bold marks the best value in each column.
Figure 6: Visual examples of hue-rotated inputs. A single ImageNet validation image is shown after rotating its hue from 0∘0^{\circ} to 330∘330^{\circ} in 30∘30^{\circ} steps (12 images). The label HH denotes the relative hue value in HSV space ( H∈[0,1)H\in[0,1) , with 11 corresponding to 360∘360^{\circ} ).
Figure 7: Hue-rotation robustness across all 9 models. a) Top-1 accuracy at each hue angle ( 0∘0^{\circ} – 324∘324^{\circ} in 36∘36^{\circ} steps). b) Accuracy drop relative to 0∘0^{\circ} .
 ImageNet-valgrayscalesketchB&Wcannysilhouette
baseline73.9062.4523.7041.3910.833.30
ED_Student69.7064.6428.0849.0113.275.43
ED_SoftLabelOnly53.5052.5824.4734.8711.455.42
Self Distilled74.7262.9724.4342.6911.593.24
Gray Distilled74.6069.5125.6045.567.873.63
B&W Distilled64.8156.2023.4745.5412.123.22
Appendix
Table 14: Top-1 accuracy (%) on the ImageNet validation set with color/texture information progressively stripped. Bold marks the highest in each column, underline the second. Referenced from both Appendix B.2 and C.2 . Visual examples in Figure 8 .
Figure 8: Visual examples of each transformation. a) RGB, b) grayscale, c) sketch, d) B&W (Otsu thresholding), e) Canny, f) silhouette, g) synthetic event. The last panel is the two-channel ON/OFF map used to train the Synth_Distilled teacher (Section A.8 ), rendered with ON in red and OFF in blue; it is the grayscale difference between two augmented views of a) and is not the output of an event simulator or of a real event camera.
Figure 9: Feature maps in the early layers (Layer 0–2) of the baseline model. Each column corresponds to a hue rotation angle ( 0∘0^{\circ} – 330∘330^{\circ} in 30∘30^{\circ} steps), and each row corresponds to one channel. Red boxes mark channels whose response varies substantially with hue change (not color-invariant). See Figure 6 for visual examples of the hue-rotated inputs.
Figure 10: Early-layer feature maps of ED_SoftLabelOnly. Same format as Figure 9 . Responses are notably invariant to hue rotation; only a few channels are flagged in red.
Figure 11: Cue-conflict experiments across architectures and models. a) Shape-bias matrixplot of all 9 ResNet-34 models (baseline, ED_Student, ED_SoftLabelOnly, Self/Gray/B&W Distilled, Adversarial Training, Noise Training, SIN) and the human reference, computed using the bethgelab/model-vs-human 16-class restricted-argmax protocol [ 1 ] . b) VGG-16 and c) ConvNeXt-Base, each with the three models baseline / ED_Student / ED_SoftLabelOnly. Each panel plots the texture-versus-shape decision ratio per category as a scatter plot.
Figure 12: Patch-shuffle / patch-rotation experiments. a) Visual examples. b) Top-1 accuracy under patch shuffle as a function of grid granularity (baseline vs. Event-distilled models). c) Accuracy under patch shuffle for all 9 models. d) Same for patch rotation.
Figure 13: Cross-architecture comparison of patch-shuffle / patch-rotation experiments. For VGG-16, ConvNeXt-Base, ResNet-50, and RepLKNet31B, we report the top-1 accuracy progression of baseline / ED_Student / ED_SoftLabelOnly as the number of patch divisions varies. Top: patch shuffle. Bottom: patch rotation.
Figure 14: All 64 channels of the first convolutional layer of ResNet-34. a) baseline, b) ED_Student, c) ED_SoftLabelOnly. Whereas Figure 3 (a) shows excerpted channels, this figure shows the layer-wide tendency.
Correlation TypeRGB CorrelationSpatial Autocorrelation
baseline0.5490.391
ED_Student0.5440.398
ED_SoftLabelOnly0.7470.575
Appendix
Table 15: Correlation values for the 64 channels of the first convolutional layer of ResNet-34. RGB Correlation is the cross-channel correlation, and Spatial Autocorrelation is between adjacent pixels. Higher values indicate smoother, more isotropic filter structures.
Figure 15: Robustness to frequency-band random noise. a) Examples of inputs perturbed by band-limited noise ( left : center frequency 0.15π0.15\pi , right : 0.95π0.95\pi ). b) Top-1 accuracy (absolute) under noise concentrated at each band; horizontal axis is the center frequency of the added noise ( 00 – π\pi ).
Figure 16: Top-1 accuracy of ResNet-34 models (baseline, ED_Student, ED_SoftLabelOnly) against all 19 ImageNet-C corruptions, grouped into Blur / Digital / Noise / Weather. Horizontal axis is Severity ( 11 – 55 ).
Figure 17: An ImageNet validation image with additive Gaussian Noise (Full range, [0,π][0,\pi] ), and the results of first decomposing the noise with band-pass filters into Low ( [0,0.1π][0,0.1\pi] ), Mid ( [0.1π,0.3π][0.1\pi,0.3\pi] ), High ( [0.3π,0.6π][0.3\pi,0.6\pi] ), Very high ( [0.6π,0.85π][0.6\pi,0.85\pi] ), and Extreme ( [0.85π,π][0.85\pi,\pi] ) and then adding each band’s component to the image. The Full-range noise visually appears dominated by high-frequency granular components, but it carries energy across all frequency bands.
Figure 18: The same ImageNet image (RGB, top row) and its Canny edge map (Canny edges, bottom row) decomposed by band-pass filters into the same bands as in Figure 17 . The Original column shows the input that each row decomposes (top: the RGB image; bottom: the Canny-processed image). The energy of the RGB image (top) is concentrated in the Low–Mid bands ( [0,0.3π][0,0.3\pi] ). Even the Canny map (bottom), which consists solely of edge components, similarly concentrates its energy in the Low–Mid bands and carries virtually no energy in the Very-high or higher bands. In other words, edge information is not localized in the very-high band: the low-frequency band still carries broad edge components accompanied by gradual gradients, and the energy is biased toward the mid-frequency band.
Figure 19: Radial PSD of ImageNet (natural images) and N-ImageNet (event representation) plotted in log-log space. Linear fits are computed over the intermediate frequency band ( 0.02F≤f≤0.25F0.02F\leq f\leq 0.25F , gray shading; FF denotes the maximum frequency) to estimate the power-law exponent α\alpha in P(f)∝1/fαP(f)\propto 1/f^{\alpha} . ImageNet yields α=2.46\alpha=2.46 ( R2=0.9993R^{2}=0.9993 ), broadly consistent with the empirically established 1/f21/f^{2} statistics of natural images [ 99 ] , while N-ImageNet yields α=1.48\alpha=1.48 ( R2=0.9996R^{2}=0.9996 ) — a markedly shallower slope, indicating less concentration of energy at low frequencies.
More in this topic

Event-Based Vision

Event-based vision uses asynchronous camera measurements to perceive motion and scene changes with high temporal resolution and dynamic range.

See all 236 papers

Explore similar work

CardsList

  1. Sensor Generalization for Adaptive Sensing in Event-based Object Detection via Joint Distribution Training

    Feb 26, 2026Aheli Saha, René Schuster, Didier StrickerAdaptive SensingEvent-Based Vision

    Deutsches Forschungszentrum für Künstliche Intelligenz, Kaiserslautern, Germany

  2. Hyper-RED: Scalable Event Pre-training via Semantic Hypergraph Distillation

    Sep 15, 2026Meisen Wang, Zhiqiang Tian, Wei Bao +3Event-Based VisionCross-Modal Knowledge Distillation

    School of Software Engineering, Xi’an Jiaotong University, Xi’an 710049, China · National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, Xi’an 710049, China · BNRist, THUIBCS, BLBCI, School of Software, Tsinghua University, Beijing 100084, China +2

  3. REALM: An RGB- and Event-Aligned Latent Manifold for Cross-Modal Perception

    Apr 30, 2026Vincenzo Polizzi, David B. Lindell, Jonathan KellyMultimodal Foundation ModelsEvent-Based Vision

    University of Toronto, Robotics Institute · University of Toronto, Department of Computer Science