Datasets:

atharvadagaonkar
/

ImageNet-CJ

file_name stringlengths 22 57	image_id stringclasses 500 values	arm stringclasses 4 values	corruption stringclasses 16 values	jpeg_quality float64 75 90 ⌀	wnid stringclasses 10 values	imagenet_index int64 0 701
data/base/img_0001.png	img_0001	clean_png	clean	null	n01440764	0
data/base/img_0002.png	img_0002	clean_png	clean	null	n01440764	0
data/base/img_0003.png	img_0003	clean_png	clean	null	n01440764	0
data/base/img_0004.png	img_0004	clean_png	clean	null	n01440764	0
data/base/img_0005.png	img_0005	clean_png	clean	null	n01440764	0
data/base/img_0006.png	img_0006	clean_png	clean	null	n01440764	0
data/base/img_0007.png	img_0007	clean_png	clean	null	n01440764	0
data/base/img_0008.png	img_0008	clean_png	clean	null	n01440764	0
data/base/img_0009.png	img_0009	clean_png	clean	null	n01440764	0
data/base/img_0010.png	img_0010	clean_png	clean	null	n01440764	0
data/base/img_0011.png	img_0011	clean_png	clean	null	n01440764	0
data/base/img_0012.png	img_0012	clean_png	clean	null	n01440764	0
data/base/img_0013.png	img_0013	clean_png	clean	null	n01440764	0
data/base/img_0014.png	img_0014	clean_png	clean	null	n01440764	0
data/base/img_0015.png	img_0015	clean_png	clean	null	n01440764	0
data/base/img_0016.png	img_0016	clean_png	clean	null	n01440764	0
data/base/img_0017.png	img_0017	clean_png	clean	null	n01440764	0
data/base/img_0018.png	img_0018	clean_png	clean	null	n01440764	0
data/base/img_0019.png	img_0019	clean_png	clean	null	n01440764	0
data/base/img_0020.png	img_0020	clean_png	clean	null	n01440764	0
data/base/img_0021.png	img_0021	clean_png	clean	null	n01440764	0
data/base/img_0022.png	img_0022	clean_png	clean	null	n01440764	0
data/base/img_0023.png	img_0023	clean_png	clean	null	n01440764	0
data/base/img_0024.png	img_0024	clean_png	clean	null	n01440764	0
data/base/img_0025.png	img_0025	clean_png	clean	null	n01440764	0
data/base/img_0026.png	img_0026	clean_png	clean	null	n01440764	0
data/base/img_0027.png	img_0027	clean_png	clean	null	n01440764	0
data/base/img_0028.png	img_0028	clean_png	clean	null	n01440764	0
data/base/img_0029.png	img_0029	clean_png	clean	null	n01440764	0
data/base/img_0030.png	img_0030	clean_png	clean	null	n01440764	0
data/base/img_0031.png	img_0031	clean_png	clean	null	n01440764	0
data/base/img_0032.png	img_0032	clean_png	clean	null	n01440764	0
data/base/img_0033.png	img_0033	clean_png	clean	null	n01440764	0
data/base/img_0034.png	img_0034	clean_png	clean	null	n01440764	0
data/base/img_0035.png	img_0035	clean_png	clean	null	n01440764	0
data/base/img_0036.png	img_0036	clean_png	clean	null	n01440764	0
data/base/img_0037.png	img_0037	clean_png	clean	null	n01440764	0
data/base/img_0038.png	img_0038	clean_png	clean	null	n01440764	0
data/base/img_0039.png	img_0039	clean_png	clean	null	n01440764	0
data/base/img_0040.png	img_0040	clean_png	clean	null	n01440764	0
data/base/img_0041.png	img_0041	clean_png	clean	null	n01440764	0
data/base/img_0042.png	img_0042	clean_png	clean	null	n01440764	0
data/base/img_0043.png	img_0043	clean_png	clean	null	n01440764	0
data/base/img_0044.png	img_0044	clean_png	clean	null	n01440764	0
data/base/img_0045.png	img_0045	clean_png	clean	null	n01440764	0
data/base/img_0046.png	img_0046	clean_png	clean	null	n01440764	0
data/base/img_0047.png	img_0047	clean_png	clean	null	n01440764	0
data/base/img_0048.png	img_0048	clean_png	clean	null	n01440764	0
data/base/img_0049.png	img_0049	clean_png	clean	null	n01440764	0
data/base/img_0050.png	img_0050	clean_png	clean	null	n01440764	0
data/base/img_0051.png	img_0051	clean_png	clean	null	n02102040	217
data/base/img_0052.png	img_0052	clean_png	clean	null	n02102040	217
data/base/img_0053.png	img_0053	clean_png	clean	null	n02102040	217
data/base/img_0054.png	img_0054	clean_png	clean	null	n02102040	217
data/base/img_0055.png	img_0055	clean_png	clean	null	n02102040	217
data/base/img_0056.png	img_0056	clean_png	clean	null	n02102040	217
data/base/img_0057.png	img_0057	clean_png	clean	null	n02102040	217
data/base/img_0058.png	img_0058	clean_png	clean	null	n02102040	217
data/base/img_0059.png	img_0059	clean_png	clean	null	n02102040	217
data/base/img_0060.png	img_0060	clean_png	clean	null	n02102040	217
data/base/img_0061.png	img_0061	clean_png	clean	null	n02102040	217
data/base/img_0062.png	img_0062	clean_png	clean	null	n02102040	217
data/base/img_0063.png	img_0063	clean_png	clean	null	n02102040	217
data/base/img_0064.png	img_0064	clean_png	clean	null	n02102040	217
data/base/img_0065.png	img_0065	clean_png	clean	null	n02102040	217
data/base/img_0066.png	img_0066	clean_png	clean	null	n02102040	217
data/base/img_0067.png	img_0067	clean_png	clean	null	n02102040	217
data/base/img_0068.png	img_0068	clean_png	clean	null	n02102040	217
data/base/img_0069.png	img_0069	clean_png	clean	null	n02102040	217
data/base/img_0070.png	img_0070	clean_png	clean	null	n02102040	217
data/base/img_0071.png	img_0071	clean_png	clean	null	n02102040	217
data/base/img_0072.png	img_0072	clean_png	clean	null	n02102040	217
data/base/img_0073.png	img_0073	clean_png	clean	null	n02102040	217
data/base/img_0074.png	img_0074	clean_png	clean	null	n02102040	217
data/base/img_0075.png	img_0075	clean_png	clean	null	n02102040	217
data/base/img_0076.png	img_0076	clean_png	clean	null	n02102040	217
data/base/img_0077.png	img_0077	clean_png	clean	null	n02102040	217
data/base/img_0078.png	img_0078	clean_png	clean	null	n02102040	217
data/base/img_0079.png	img_0079	clean_png	clean	null	n02102040	217
data/base/img_0080.png	img_0080	clean_png	clean	null	n02102040	217
data/base/img_0081.png	img_0081	clean_png	clean	null	n02102040	217
data/base/img_0082.png	img_0082	clean_png	clean	null	n02102040	217
data/base/img_0083.png	img_0083	clean_png	clean	null	n02102040	217
data/base/img_0084.png	img_0084	clean_png	clean	null	n02102040	217
data/base/img_0085.png	img_0085	clean_png	clean	null	n02102040	217
data/base/img_0086.png	img_0086	clean_png	clean	null	n02102040	217
data/base/img_0087.png	img_0087	clean_png	clean	null	n02102040	217
data/base/img_0088.png	img_0088	clean_png	clean	null	n02102040	217
data/base/img_0089.png	img_0089	clean_png	clean	null	n02102040	217
data/base/img_0090.png	img_0090	clean_png	clean	null	n02102040	217
data/base/img_0091.png	img_0091	clean_png	clean	null	n02102040	217
data/base/img_0092.png	img_0092	clean_png	clean	null	n02102040	217
data/base/img_0093.png	img_0093	clean_png	clean	null	n02102040	217
data/base/img_0094.png	img_0094	clean_png	clean	null	n02102040	217
data/base/img_0095.png	img_0095	clean_png	clean	null	n02102040	217
data/base/img_0096.png	img_0096	clean_png	clean	null	n02102040	217
data/base/img_0097.png	img_0097	clean_png	clean	null	n02102040	217
data/base/img_0098.png	img_0098	clean_png	clean	null	n02102040	217
data/base/img_0099.png	img_0099	clean_png	clean	null	n02102040	217
data/base/img_0100.png	img_0100	clean_png	clean	null	n02102040	217

End of preview. Expand in Data Studio

JPEG Re-encoding Confound Control Dataset

A controlled-experiment dataset that isolates one acknowledged-but-unmeasured confound in ImageNet-C. Hendrycks & Dietterich (Benchmarking Neural Network Robustness to Common Corruptions and Perturbations, ICLR 2019, arXiv:1903.12261) save every corrupted image as a lightly compressed JPEG. The benchmark therefore never measures a corruption c applied to an image x in isolation — it measures JPEG(c(x)). This dataset lets you quantify how much of each corruption's reported error is attributable to the corruption versus the final JPEG save step, and whether that leakage differs across corruption families.

The single isolated variable is the save format: the same corrupted pixel array is saved once losslessly (PNG) and once as JPEG. Everything else — image, corruption, severity, random seed — is held identical, so any measured difference is attributable to the JPEG save step and nothing else.

The control structure (four matched arms)

For each base image, reduced once to a canonical 224×224 RGB array:

arm	description
`clean_png`	canonical image saved losslessly as PNG — the control arm
`clean_jpeg`	canonical image saved as JPEG at quality `q` — isolates JPEG acting alone
`corrupt_png`	corrupted array saved losslessly as PNG — the corruption in isolation
`corrupt_jpeg`	the same corrupted array saved as JPEG at quality `q` — what ImageNet-C stores

The corruption array for a given (image, corruption, severity) is generated exactly once and then saved every way; it is never regenerated per format. That one-array rule is the core of the experiment's validity. The corruption functions are the original ImageNet-C code, used via the imagecorruptions package — the only logic this project adds is Image.save(format='PNG') vs Image.save(format='JPEG', quality=q).

Scope

15 corruptions (families: noise, blur, weather, digital).
3 severities: {1, 3, 5}.
3 JPEG qualities: {75, 85, 90}. Quality 85 is the ImageNet-C default, read directly from make_imagenet_c.py in the Hendrycks repo (save(..., quality=85, optimize=True)); 75 and 90 bracket it so results are not tied to one guess.
Base images: the 10-class ImageNette (320px) validation split, selected deterministically from a fixed seed.

Files and layout

data/
  base/<image_id>.png                                # clean_png arm
  clean_saves/jpeg_q<q>/<image_id>.jpg               # clean_jpeg arm
  corrupted/<corruption>/s<severity>/
      png/<image_id>.png                             # corrupt_png arm
      jpeg_q<q>/<image_id>.jpg                        # corrupt_jpeg arm
manifest.csv     # one row per base image: image_id, source_path, wnid, imagenet_index, base_png_path
metadata.csv     # one row per arm file (index over the whole dataset; schema below)
config.yaml      # exact scope knobs used to generate this release

<image_id> (e.g. img_0001) is assigned deterministically at selection time. The base/ PNG doubles as the clean_png arm and is not duplicated.

`metadata.csv` schema

One row per image file, so the dataset can be filtered without walking the tree:

column	meaning
`file_name`	path relative to repo root, e.g. `data/corrupted/fog/s3/jpeg_q85/img_0001.jpg`
`image_id`	stable id, e.g. `img_0001`
`arm`	one of `clean_png`, `clean_jpeg`, `corrupt_png`, `corrupt_jpeg`
`corruption`	corruption name, or `clean` for the clean arms
`severity`	`1/3/5`, or `0` for clean arms
`jpeg_quality`	`75/85/90` for JPEG arms, empty for PNG arms
`wnid`	ImageNet wnid of the source class
`imagenet_index`	ImageNet-1k class index (0–999)

Usage

Pair each *_png arm with the matching *_jpeg arm (same image_id, corruption, severity) and compare model behavior. The intended contrast is corrupt_png vs corrupt_jpeg; the clean_* arms isolate JPEG acting alone (the control).

import pandas as pd
from huggingface_hub import hf_hub_download

repo = "<user>/jpeg-confound-control-dataset"
meta = pd.read_csv(hf_hub_download(repo, "metadata.csv", repo_type="dataset"))

# all matched arms for fog at severity 3, quality 85
fog = meta[(meta.corruption == "fog") & (meta.severity == 3) &
           (meta.arm.isin(["corrupt_png", "corrupt_jpeg"])) &
           (meta.jpeg_quality.isin([85, ""]))]

Or pull the full tree:

from huggingface_hub import snapshot_download
snapshot_download("<user>/jpeg-confound-control-dataset", repo_type="dataset")

Reproducibility

The dataset regenerates byte-identically from the generation code plus the seed in config.yaml: image selection is seeded, and before every corruption an order-independent 32-bit seed is derived from (image_id, corruption, severity) via zlib.crc32. The full generator, evaluator, and analysis are in the code repository (generate.py → evaluate.py → analyze.py).

Intended use and limitations

Intended use: measuring the JPEG save-step confound in corruption-robustness benchmarks; teaching controlled-experiment design. Research only.
No lossless ImageNet original exists — ImageNet (and thus ImageNette) source images are themselves JPEG. This source compression is held identical across the PNG and JPEG arms, so it cancels in the comparison; the variable isolated here is the final save-time re-encode, not total JPEG exposure.
Base images are the 10-class ImageNette subset with a single canonical crop, so absolute accuracies differ from the full ImageNet-C protocol; the controlled contrast between arms is unaffected.

Licensing and source

Images derive from ImageNette, a subset of ImageNet, and are subject to the ImageNet terms of access (non-commercial research use). Redistribute only as permitted by those terms; when in doubt, release code + manifest + a small sample and let users regenerate the full set.
Corruption functions: imagecorruptions (Apache-2.0), packaging Hendrycks's original ImageNet-C code.
Packaging/generation code: see the code repository's license.

Citation

The corruptions and the confound this dataset measures:

@inproceedings{hendrycks2019benchmarking,
  title     = {Benchmarking Neural Network Robustness to Common Corruptions and Perturbations},
  author    = {Hendrycks, Dan and Dietterich, Thomas},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2019},
  url       = {https://arxiv.org/abs/1903.12261}
}