Cutout
Preprocessing Methods AIMech:Cutout DRAFT
A training-image augmentation method that removes information from a randomly located square region by replacing that region with zeros in the normalized input while keeping the original target.
Identity Notes & Concept Scope
This record denotes the original random square masking procedure. It is distinct from unit-level Dropout, input-and-label MixUp, replacement with content from another image, and an augmentation policy that merely includes a cutout operation. A policy implementation can use a different fill value or mask-size convention without redefining the original paper’s exact rule. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 238 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in both worktree and main checkout, returning 34 lines in each from prior discovery, saved reviews and related native records. Ten exact AIO CLASS nodes, their scoped synonyms and outgoing edges, plus five neighboring native mechanisms were inspected. This absence assessment is bounded to the pinned snapshot and inspected semantic neighborhood.
Claims & Grounded Evidence
1 Claim
Mechanism
#method
For each training image, randomly choose a pixel within the image as the center of a fixed-size square mask. Set pixels in the overlapping image region to zero; the square may extend beyond the image boundaries, so its visible area can be smaller than the specified square. Apply the mask during training with the original label. The paper applies cutout to zero-centered data: zero denotes the normalization origin rather than necessarily a black pixel in the unnormalized image. Mask size is a tunable parameter. At test time, use the complete image without cutout or the activation rescaling associated with dropout. The earlier feature-activation-guided masking investigation and the fully-contained mask used with 50% probability are explored alternatives, not the final random-center rule.
⚠️ Curated Limitations & Negative Results
- A removed region can contain the class-defining information; suitable mask size depends on image scale, data and task.
- Masking with unchanged labels assumes that the remaining image supports the original target. This is an inductive bias, not a semantic guarantee.
- The record does not assert a universal mask length, implementation-specific integer rounding, channel convention, benchmark gain or robustness guarantee.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:DataAugmentation
Data Augmentation
|
broad_match | Cutout is one particular transformation of training images that creates masked examples. The pinned AIO class denotes the broader data-augmentation method category; the relation is local-to-broader-target and does not edit the ontology. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Terrance DeVries, Graham W. Taylor
|
primary paper
full_text
|
ID: arXiv:1708.04552v2
sha256:ae7e175b3b93…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T10:34:59Z
Updated: 2026-10-11T10:34:59Z
Source File:
data/curated/Cutout.yaml