arxiv: 2604.26806 · v1 · submitted 2026-04-29 · 💻 cs.CV · cs.AI

Recognition: unknown

ViCrop-Det: Spatial Attention Entropy Guided Cropping for Training-Free Small-Object Detection

Hui Wang , Hongze Li , Wei Chen , Xiaojin Zhang

Authors on Pith no claims yet

Pith reviewed 2026-05-07 11:03 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords small-object detectionattention entropytransformer detectortraining-freeadaptive croppingspatial trust regionDETRcross-attention

0 comments

The pith

ViCrop-Det improves small-object detection by using cross-attention entropy to guide adaptive cropping of high-ambiguity regions in transformer detectors.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Transformer detectors suffer feature degradation on small objects because they impose one uniform receptive field over images that contain regions of very different information density. ViCrop-Det extracts Spatial Attention Entropy from the decoder cross-attention maps to locate the zones of greatest local uncertainty and target saliency. It then shrinks the spatial trust region to those zones alone and injects higher-resolution observations there. The procedure runs without any retraining or model changes and produces 1-3 point gains in mAP@50 on VisDrone and DOTA-v1.5 while adding only 20-23 percent latency. On COCO the same routine raises small-object AP while leaving medium- and large-object AP unchanged.

Core claim

By treating the detection decoder's cross-attention distribution as an endogenous probe, ViCrop-Det computes Spatial Attention Entropy to measure local spatial ambiguity, then performs dynamic spatial routing that allocates a fixed computational budget exclusively to regions of both high target saliency and high uncertainty. Shrinking the spatial trust region and inserting high-frequency localized observations resolves the fine-grained degradation that uniform global receptive fields cause for microscopic targets.

What carries the argument

Spatial Attention Entropy (SAE) computed from cross-attention maps, used as a heuristic to select and shrink spatial trust regions for focused high-frequency observation.

If this is right

Adds +1-3 mAP@50 to RT-DETR-R50 and Deformable DETR on VisDrone and DOTA-v1.5 at 20-23 percent extra latency.
Raises AP_S on MS COCO while AP_M and AP_L stay stable.
Outperforms uniform slicing baselines when total compute is held constant.
Requires no architectural changes or additional training data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same entropy probe could be applied inside other transformer vision pipelines that currently rely on fixed multi-scale pyramids.
Because the method preserves the original global prior outside the cropped zones, it may reduce the need for separate small-object heads in future detectors.
The routing logic might transfer directly to video object detection where small targets move between frames of varying clarity.

Load-bearing premise

Cross-attention entropy maps reliably mark the precise locations of small objects or high-ambiguity zones whose cropping recovers useful features without missing other objects or creating new errors.

What would settle it

A set of images containing verified small objects in which the SAE maps assign low entropy to those objects; applying the cropping step on those images produces no gain or a measurable drop in detection recall.

Figures

Figures reproduced from arXiv: 2604.26806 by Hongze Li, Hui Wang, Wei Chen, Xiaojin Zhang.

**Figure 1.** Figure 1: ViCrop-Det processing pipeline. and calculate the normalized Shannon entropy: Hˆ s = − 1 log Nk X i∈P pi log pi (3) where Nk is the total number of spatial keys. In this formulation, Hˆ s ∈ [0, 1] serves as an effective, endogenous indicator of the model’s spatial cognitive ambiguity. A highly dispersed attention map yields a high Hˆ s , indicating that the local feature representation is severely entang… view at source ↗

**Figure 2.** Figure 2: The RT-DETR-R50 model applies ViCrop-Det e view at source ↗

**Figure 3.** Figure 3: The RT-DETR-R50 model applies ViCrop-Det e view at source ↗

read the original abstract

Transformer-based architectures have established a dominant paradigm in global semantic perception; however, they remain fundamentally constrained by the profound spatial heterogeneity inherent in natural images. Specifically, the imposition of a uniform global receptive field across regions of varying information density inevitably leads to local feature degradation, particularly in dense conflict zones populated by microscopic targets. To address this mechanistic limitation, we propose ViCrop-Det, a training-free inference framework that introduces adaptive spatial trust region shrinkage. Inspired by the use of attention entropy in anomaly segmentation, ViCrop-Det leverages the detection decoder's cross-attention distribution as an endogenous probe. By utilizing Spatial Attention Entropy (SAE) to heuristically evaluate local spatial ambiguity, the framework executes dynamic spatial routing, allocating a fixed computational budget exclusively to regions exhibiting both high target saliency and high cognitive uncertainty. By shrinking the spatial trust region and injecting high-frequency localized observations, ViCrop-Det actively resolves spatial ambiguity and recovers fine-grained features without requiring architectural modifications. Extensive evaluations on VisDrone and DOTA-v1.5 demonstrate that ViCrop-Det yields competitive performance enhancements, consistently adding +1-3 mAP@50 to RT-DETR-R50 and Deformable DETR with a marginal 20-23\% latency overhead. On MS COCO, $AP_{S}$ improves while $AP_{M}/AP_{L}$ remains stable, indicating precise fine-scale refinement without compromising the global spatial prior. Under compute-matched settings, our adaptive routing strategy comprehensively surpasses uniform slicing baselines, achieving a highly optimized accuracy-speed trade-off.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ViCrop-Det's entropy-guided cropping is a sensible inference-only tweak for small objects in DETR models, but the +1-3 mAP claims rest on thin evidence and an untested assumption about initial attention.

read the letter

ViCrop-Det uses the entropy of cross-attention maps from the decoder to pick regions for cropping and higher-resolution processing at inference time. The goal is to recover fine details on small targets without retraining or changing the model architecture. That specific routing step based on spatial attention entropy stands out as the actual new piece; it is not just another uniform slice or standard attention visualization trick. The paper applies it to RT-DETR and Deformable DETR on VisDrone, DOTA, and COCO, reporting consistent small-object lifts with modest latency cost. If the numbers hold, the method could be a practical add-on for aerial or dense-scene pipelines where global receptive fields lose tiny objects. The core logic is straightforward and targets a known weakness in transformer detectors. The main weakness is that the headline gains (+1-3 mAP@50, 20-23% latency) are stated without ablations, error bars, or detailed baseline tables in the available text. More importantly, the approach assumes the first forward pass already produces usable attention on the small objects worth refining. When that assumption fails, the entropy map will not highlight the right areas, and the cropping step adds overhead without fixing the miss. The stress-test concern is real and not obviously mitigated by the training-free design. The paper is aimed at engineers and researchers who run DETR-style detectors on imagery with lots of small targets and limited compute. A reader already working on inference optimizations could extract the routing idea and test it themselves. It deserves peer review because the technique is distinct enough to warrant proper experimental scrutiny and feedback on whether the circularity issue was handled in the full results.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes ViCrop-Det, a training-free inference framework for small-object detection in transformer detectors (RT-DETR-R50, Deformable DETR). It computes Spatial Attention Entropy (SAE) from the decoder's cross-attention distribution as an endogenous probe to identify regions of high target saliency and spatial ambiguity, then applies adaptive spatial cropping to allocate a fixed compute budget for high-resolution refinement on those regions. The central claims are consistent +1-3 mAP@50 gains on VisDrone and DOTA-v1.5 with 20-23% latency overhead, plus improved AP_S on MS COCO while preserving AP_M/AP_L, and superiority over uniform slicing under compute-matched conditions.

Significance. If the reported gains prove reproducible with proper controls, the work would provide a practical, architecture-agnostic way to mitigate spatial heterogeneity and local feature degradation in global-receptive-field transformers without retraining. The training-free design and modest overhead are genuine strengths that could enable immediate deployment on existing models. The heuristic use of attention entropy for dynamic routing is intuitive and extends prior ideas from anomaly segmentation, though its load-bearing assumptions require stronger validation.

major comments (3)

[Abstract] Abstract: the specific numeric claims (+1-3 mAP@50, 20-23% latency overhead, AP_S improvement) are stated without any implementation details, baseline descriptions, error bars, ablation tables, or statistical significance tests. This directly undermines verifiability of the central empirical result.
[Method] Method description (cross-attention entropy routing): the assumption that SAE computed from the initial decoder cross-attention on the full image reliably surfaces small-object locations is not tested against failure modes of the base detector. When the first-pass attention on tiny targets is weak or absent, the entropy map cannot flag those regions, so the claimed refinement gains cannot materialize; no targeted experiments or failure-case analysis address this circularity between probe and refinement.
[Results] Results section: the statement that the adaptive strategy 'comprehensively surpasses uniform slicing baselines' under compute-matched settings lacks any description of how the compute budget is equalized, the exact number of crops, or quantitative tables with variance; without these, the accuracy-speed trade-off claim cannot be assessed.

minor comments (2)

[Abstract] The acronym SAE is introduced in the abstract without an immediate parenthetical definition or pointer to the equation that defines it.
[Abstract] Notation for mAP@50 and AP_S should be explicitly defined on first use even if standard in the field.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below, indicating where we will revise the manuscript to improve clarity, verifiability, and rigor while preserving the core contributions.

read point-by-point responses

Referee: [Abstract] Abstract: the specific numeric claims (+1-3 mAP@50, 20-23% latency overhead, AP_S improvement) are stated without any implementation details, baseline descriptions, error bars, ablation tables, or statistical significance tests. This directly undermines verifiability of the central empirical result.

Authors: We agree that the abstract would benefit from greater self-containment to support immediate verifiability. In the revised version we will expand the abstract to explicitly name the base detectors (RT-DETR-R50, Deformable DETR), the three evaluation datasets, and the fact that all reported gains are accompanied by full ablation tables, standard deviations, and compute-matched baselines in Sections 4 and 5. Because abstracts are length-constrained, exhaustive error-bar tables and significance tests will remain in the main body, but the claims will be better contextualized. revision: partial
Referee: [Method] Method description (cross-attention entropy routing): the assumption that SAE computed from the initial decoder cross-attention on the full image reliably surfaces small-object locations is not tested against failure modes of the base detector. When the first-pass attention on tiny targets is weak or absent, the entropy map cannot flag those regions, so the claimed refinement gains cannot materialize; no targeted experiments or failure-case analysis address this circularity between probe and refinement.

Authors: This is a substantive and fair concern about the probe's reliability. The SAE heuristic presupposes that the base decoder's initial cross-attention map contains usable signal for small objects; complete misses would indeed leave those regions unflagged. While the manuscript includes qualitative attention visualizations (Figure 3) showing successful localization of small targets, we did not provide a dedicated failure-mode study or quantitative analysis of cases where base attention is near-zero. In the revision we will add a short subsection in the Method discussion that explicitly acknowledges this assumption, its boundary conditions, and the resulting limitation, together with additional qualitative examples of weak-attention cases. revision: yes
Referee: [Results] Results section: the statement that the adaptive strategy 'comprehensively surpasses uniform slicing baselines' under compute-matched settings lacks any description of how the compute budget is equalized, the exact number of crops, or quantitative tables with variance; without these, the accuracy-speed trade-off claim cannot be assessed.

Authors: We accept that the compute-matching protocol was described too briefly. In the revised Results section we will insert a dedicated paragraph explaining the equalization procedure (matching total wall-clock latency and approximate FLOPs), state the precise number of uniform crops used for each baseline, and augment the relevant tables with mean and standard-deviation values computed over multiple runs. These additions will allow direct assessment of the accuracy-speed trade-off. revision: yes

Circularity Check

0 steps flagged

No derivation chain; heuristic framework is self-contained

full rationale

The paper describes ViCrop-Det as a training-free inference-time heuristic that computes Spatial Attention Entropy (SAE) from an existing detector's cross-attention map and uses it to route cropping. No equations, fitted parameters, or formal derivations appear in the provided text. The central mechanism is presented as an empirical routing strategy inspired by prior anomaly-segmentation work, not as a prediction derived from or equivalent to its own inputs. No self-citations are shown to be load-bearing, no uniqueness theorems are invoked, and no ansatz or renaming reduces the claimed gains to a tautology. The method therefore contains no circular reduction of the sort enumerated in the analysis criteria.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the unproven correlation between attention entropy and object ambiguity, drawn from prior anomaly segmentation literature, plus the assumption that focused cropping improves fine-scale features.

axioms (1)

domain assumption Cross-attention distributions from the detection decoder serve as an endogenous probe for local spatial ambiguity
Invoked to justify SAE as the routing signal; no independent validation supplied in the abstract.

invented entities (1)

Spatial Attention Entropy (SAE) no independent evidence
purpose: Heuristic measure of local spatial ambiguity to guide cropping decisions
Introduced as the key endogenous signal; independent evidence outside the paper is not provided.

pith-pipeline@v0.9.0 · 5587 in / 1345 out tokens · 80372 ms · 2026-05-07T11:03:09.814063+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

32 extracted references · 3 canonical work pages · 1 internal anchor

[1]

Attention is all you need.Advances in neural information pro- cessing systems, 30, 2017

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information pro- cessing systems, 30, 2017

2017
[2]

End-to-end object detection with transformers

Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuropean conference on computer vision, pages 213–229. Springer, 2020

2020
[3]

A mathematical theory of communica- tion.The Bell system technical journal, 27(3):379–423, 1948

Claude Elwood Shannon. A mathematical theory of communica- tion.The Bell system technical journal, 27(3):379–423, 1948

1948
[4]

Slicing aided hyper inference and fine-tuning for small object detection

Fatih Cagatay Akyon, Sinan Onur Altinuc, and Alptekin Temizel. Slicing aided hyper inference and fine-tuning for small object detection. In2022 IEEE international conference on image pro- cessing (ICIP), pages 966–970. IEEE, 2022

2022
[5]

Attentropy: On the generalization ability of supervised semantic segmentation transformers to new objects in new domains

Krzysztof Baron-Lis, Matthias Rottmann, Annika Mütze, Sina Honari, Pascal Fua, and Mathieu Salzmann. Attentropy: On the generalization ability of supervised semantic segmentation transformers to new objects in new domains. In35th British Machine Vision Conference 2024, BMVC 2024, Glasgow, UK, November 25-28, 2024. BMV A, 2024. URLhttps://papers. bmvc2024....

2024
[6]

Visdrone-det2018: The vision meets drone object detec- tion in image challenge results

Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Haibin Ling, Qinghua Hu, Qinqin Nie, Hao Cheng, Chenfeng Liu, Xiaoyu Liu, et al. Visdrone-det2018: The vision meets drone object detec- tion in image challenge results. InProceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018

2018
[7]

Dota: A large-scale dataset for object detection in aerial images

Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3974–3983, 2018

2018
[8]

Detrs beat yolos on real-time object detection

Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16965–16974, 2024

2024
[9]

Deformable DETR: Deformable Transformers for End-to-End Object Detection

Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end- to-end object detection.arXiv preprint arXiv:2010.04159, 2020

work page internal anchor Pith review arXiv 2010
[10]

Adavit: Adaptive vision transformers for efficient image recognition

Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12309–12318, 2022

2022
[11]

Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition.Advances in neural information processing systems, 34:11960–11973, 2021

Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition.Advances in neural information processing systems, 34:11960–11973, 2021

2021
[12]

Dynamicvit: Efficient vision transformers with dynamic token sparsification.Advances in neural informa- tion processing systems, 34:13937–13949, 2021

Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification.Advances in neural informa- tion processing systems, 34:13937–13949, 2021

2021
[13]

Patch slimming for efficient vision transformers

Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12165–12174, 2022

2022
[14]

Tokenlearner: Adaptive space-time tokenization for videos.Advances in neural information process- ing systems, 34:12786–12797, 2021

Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa De- hghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos.Advances in neural information process- ing systems, 34:12786–12797, 2021

2021
[15]

Adaptive slicing-aided hyper inference for small object detec- tion in high-resolution remote sensing images.Remote Sensing, 15(5):1249, 2023

Hao Zhang, Chuanyan Hao, Wanru Song, Bo Jiang, and Baozhu Li. Adaptive slicing-aided hyper inference for small object detec- tion in high-resolution remote sensing images.Remote Sensing, 15(5):1249, 2023

2023
[16]

Deep interpretable classifica- tion and weakly-supervised segmentation of histology images via max-min uncertainty.IEEE Transactions on Medical Imaging, 41 (3):702–714, 2021

Soufiane Belharbi, Jérôme Rony, Jose Dolz, Ismail Ben Ayed, Luke McCaffrey, and Eric Granger. Deep interpretable classifica- tion and weakly-supervised segmentation of histology images via max-min uncertainty.IEEE Transactions on Medical Imaging, 41 (3):702–714, 2021

2021
[17]

Inter- preting cnns via decision trees

Quanshi Zhang, Yu Yang, Haotian Ma, and Ying Nian Wu. Inter- preting cnns via decision trees. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6261–6270, 2019

2019
[18]

Featup: A model- agnostic framework for features at any resolution.arXiv preprint arXiv:2403.10516, 2024

Stephanie Fu, Mark Hamilton, Laura Brandt, Axel Feldman, Zhoutong Zhang, and William T Freeman. Featup: A model- agnostic framework for features at any resolution.arXiv preprint arXiv:2403.10516, 2024

work page arXiv 2024
[19]

Learning continu- ous image representation with local implicit image function

Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continu- ous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628–8638, 2021

2021
[20]

Clustered object detection in aerial images

Fan Yang, Heng Fan, Peng Chu, Erik Blasch, and Haibin Ling. Clustered object detection in aerial images. InProceedings of the IEEE/CVF international conference on computer vision, pages 8311–8320, 2019

2019
[21]

An analysis of scale invariance in object detection snip

Bharat Singh and Larry S Davis. An analysis of scale invariance in object detection snip. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3578–3587, 2018. Preprint– ViCrop-Det 9

2018
[22]

Sniper: Efficient multi-scale training.Advances in neural information processing systems, 31, 2018

Bharat Singh, Mahyar Najibi, and Larry S Davis. Sniper: Efficient multi-scale training.Advances in neural information processing systems, 31, 2018

2018
[23]

Lsknet: A foundation lightweight backbone for remote sensing: Y

Yuxuan Li, Xiang Li, Yimain Dai, Qibin Hou, Li Liu, Yongxiang Liu, Ming-Ming Cheng, and Jian Yang. Lsknet: A foundation lightweight backbone for remote sensing: Y . li et al.International Journal of Computer Vision, 133(3):1410–1431, 2025

2025
[24]

Boundary-aware feature fusion with dual-stream attention for remote sensing small object detection

Jingnan Song, Mingliang Zhou, Jun Luo, Huayan Pu, Yong Feng, Xuekai Wei, and Weijia Jia. Boundary-aware feature fusion with dual-stream attention for remote sensing small object detection. IEEE Transactions on Geoscience and Remote Sensing, 2024

2024
[25]

Yangtao Wang, Xi Shen, Yuan Yuan, Yuming Du, Maomao Li, Shell Xu Hu, James L Crowley, and Dominique Vaufreydaz. Tokencut: Segmenting objects in images and videos with self- supervised transformer and normalized cut.IEEE transactions on pattern analysis and machine intelligence, 45(12):15790–15801, 2023

2023
[26]

arXiv preprint arXiv:2109.14279

Oriane Siméoni, Gilles Puy, Huy V V o, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels.arXiv preprint arXiv:2109.14279, 2021

work page arXiv 2021
[27]

Instances as queries

Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. InProceedings of the IEEE/CVF international conference on computer vision, pages 6910–6919, 2021

2021
[28]

Cascade r-cnn: Delving into high quality object detection

Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 6154–6162, 2018

2018
[29]

Dynamic head: Unify- ing object detection heads with attentions

Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unify- ing object detection heads with attentions. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373–7382, 2021

2021
[30]

Less is more: Focus attention for efficient detr

Dehua Zheng, Wenhui Dong, Hailin Hu, Xinghao Chen, and Yunhe Wang. Less is more: Focus attention for efficient detr. InProceedings of the IEEE/CVF international conference on computer vision, pages 6674–6683, 2023

2023
[31]

Dq-detr: Detr with dynamic query for tiny object de- tection

Yi-Xin Huang, Hou-I Liu, Hong-Han Shuai, and Wen-Huang Cheng. Dq-detr: Detr with dynamic query for tiny object de- tection. InEuropean Conference on Computer Vision, pages 290–305. Springer, 2024

2024
[32]

Microsoft coco: Common objects in context

Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEuropean confer- ence on computer vision, pages 740–755. Springer, 2014. Preprint– ViCrop-Det 10 SupplementaryMaterial Abstract This supplementary material provides addi- tional details on...

2014