RECORD
Generalization to translation shifts: a study in architectures and augmentations
Computer Vision and Pattern Recognition · Submitted 5 Jul 2022
abstract compiled · 2026-09-23 · no language-model review
The authors' main claim (c) With a sufficiently advanced augmentation pipeline ($4$ pixel crop+RandAugmentation+Erasing+MixUp), all architectures can be trained to have competitive performance in terms of in-distribution accuracy as well as generalization to large translation shifts.
The compiled record
The paper's own sentences, arranged. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
Open the compiled record
The abstract, sentence by sentence
We study how effective data augmentation is at capturing the inductive bias of carefully designed network architectures for spatial translation invariance. We evaluate various image classification architectures (antialiased, convolutional, vision transformer, and fully connected MLP networks) and data augmentation techniques towards generalization to large translation shifts. We observe that: (a) without data augmentation, all architectures, including convolutional networks with antialiased modification suffer some degradation in performance when evaluated on translated test distributions. Understandably, both the in-distribution accuracy and degradation to shifts is significantly worse for non-convolutional models. (b) The robustness of performance is improved by even a minimal augmentation of $4$ pixel random crop across all architectures. In some instances, even $1-2$ pixel random crop is sufficient. This suggests that there is a form of meta generalization from augmentation. For non-convolutional architectures, while the absolute accuracy is still low with this basic augmentation, we see substantial improvements in robustness to translation shifts. (c) With a sufficiently advanced augmentation pipeline ($4$ pixel crop+RandAugmentation+Erasing+MixUp), all architectures can be trained to have competitive performance in terms of in-distribution accuracy as well as generalization to large translation shifts.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | We study how effective data augmentation is at capturing the inductive bias of carefully designed network architectures for spatial translation invariance. | contribution 0.98 | 0.60 | 0.06 | unclear 0.68 |
| s1 | We evaluate various image classification architectures (antialiased, convolutional, vision transformer, and fully connected MLP networks) and data augmentation techniques towards generalization to large translation shifts. | method 0.58 | 0.90 | 0.05 | asserted 0.80 |
| s2 | We observe that: (a) without data augmentation, all architectures, including convolutional networks with antialiased modification suffer some degradation in performance when evaluated on translated test distributions. | result 1.00 | 0.93 | 0.09 | asserted 0.98 |
| s3 | Understandably, both the in-distribution accuracy and degradation to shifts is significantly worse for non-convolutional models. | result 0.83 | 0.77 | 0.05 | asserted 1.00 |
| s4 | (b) The robustness of performance is improved by even a minimal augmentation of $4$ pixel random crop across all architectures. | result 1.00 | 0.95 | 0.16 | asserted 1.00 |
| s5 | In some instances, even $1-2$ pixel random crop is sufficient. | result 1.00 | 0.91 | 0.05 | asserted 0.58 |
| s6 | This suggests that there is a form of meta generalization from augmentation. | interpretation 1.00 | 0.77 | 0.14 | hedged possibility 1.00 |
| s7 | For non-convolutional architectures, while the absolute accuracy is still low with this basic augmentation, we see substantial improvements in robustness to translation shifts. | result 0.97 | 0.94 | 0.08 | asserted 1.00 |
| s8 | (c) With a sufficiently advanced augmentation pipeline ($4$ pixel crop+RandAugmentation+Erasing+MixUp), all architectures can be trained to have competitive performance in terms of in-distribution accuracy as well as generalization to large translation shifts. | result 1.00 | 0.91 | 0.54 | asserted 0.93 |
What was examined, and what was not
- arXiv listing abstract v2
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
The paper
Original abstract
We study how effective data augmentation is at capturing the inductive bias of carefully designed network architectures for spatial translation invariance. We evaluate various image classification architectures (antialiased, convolutional, vision transformer, and fully connected MLP networks) and data augmentation techniques towards generalization to large translation shifts. We observe that: (a) without data augmentation, all architectures, including convolutional networks with antialiased modification suffer some degradation in performance when evaluated on translated test distributions. Understandably, both the in-distribution accuracy and degradation to shifts is significantly worse for non-convolutional models. (b) The robustness of performance is improved by even a minimal augmentation of $4$ pixel random crop across all architectures. In some instances, even $1-2$ pixel random crop is sufficient. This suggests that there is a form of meta generalization from augmentation. For non-convolutional architectures, while the absolute accuracy is still low with this basic augmentation, we see substantial improvements in robustness to translation shifts. (c) With a sufficiently advanced augmentation pipeline ($4$ pixel crop+RandAugmentation+Erasing+MixUp), all architectures can be trained to have competitive performance in terms of in-distribution accuracy as well as generalization to large translation shifts.
Human review
How the literature uses it
1 arXiv paper cite this in 3 sentences
What it cites · 15 sentences without a confident label
Uses the method or tool of 1 paper
-
In our version of advanced augmentation (AA), we use the following pipeline: first we apply (a) the basic augmentation (BA) described above, then (b) RandAugment [ref] , then (c) random erasing [ref] , and finally use (d) MixUp [ref] .
Data Augmentation by Pairing Samples for Images Classification · Augmentations · 1.00
Cites as support 1 paper
-
* resnet18_gn (11M parameters): It has been observed that batch normalization often leads to poor performance in transfer learning as the batch statistics from source task could be widely off for target task (see [ref] ).
Rethinking "Batch" in BatchNorm · Architectures · 0.97
Cites as background 1 paper
-
Most popular among these are the Vision Transformers (ViTs) and their variants [ref] .
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale · Introduction · 0.85
Discussion
Know something about this paper? Leave a comment, or highlight any passage above to comment on it.