jackstev/hw1-image-light-sources-on-off

Light Sources: On vs Off

jackstev/hw1-image-light-sources-on-off

Thirty original photographs of light sources, each labelled by whether the source is emitting, plus explicitly marked synthetic training variants. Built for CMU 24-679 (Designing and Prototyping AI Systems), Homework 1. The classification task predicts on against off from a 224x224 RGB image.

Composition

Column	Description
image	224x224 RGB image.
label	Target. 1 = source emitting, 0 = not emitting.
label_name	on or off.
source_id	Capture filename, suffixed with method and copy when synthetic.
parent_id	Source image. Equals source_id when original.
subject_id	The physical light source photographed.
augmentation	Method that produced the image, none if original.
is_augmented	False for photographs, True for synthetic variants.

Collection

Thirty photographs were taken by the author of light sources around home and campus: lamps, bulbs, fixtures and vehicle lights. Most subjects were shot twice in a row, once emitting and once not, with the camera and framing held as steady as practical so the switch state is the main difference between the two frames.

No people, faces, or personally identifying information appear in any image.

Preparation is deterministic and identical for every image and split: EXIF orientation is applied, the frame is padded to a square with neutral grey (128, 128, 128), and the result is resized to 224x224 RGB. Padding rather than cropping preserves the original aspect ratio, so no part of a scene is lost before augmentation. Labels come from the Drive folder each file sat in, never from its filename.

Splits

Split	Original images	Synthetic images	Total
train	20	320	340
validation	5	0	5
test	5	0	5

Split	Class	Images
train	off	153
train	on	187
validation	off	2
validation	on	3
test	off	2
test	on	3

Subjects were partitioned 30% to holdout, then 50% of that holdout to test. Both holdouts are entirely unaugmented.

Splitting precedes augmentation, so no variant of a held-out image can reach training. The notebook asserts that every synthetic parent belongs to the training partition.

Splitting is by photographed subject, not by image. Because most subjects were shot on and off back to back, the two frames of one lamp are near-duplicates of each other. Splitting by image could place a lamp in training and the same lamp, switched, in test, and a model would score it almost for free. Subject grouping was inferred from consecutive capture numbers and verified: every inferred pair contains exactly one on and one off image, which the notebook asserts. Six images have no partner and form single-image subjects.

The class image notebook does not group, and says so; the class text notebook does group, by survey response. This dataset follows the text notebook's reasoning because its photographs are deliberately paired.

Augmentation

Synthetic images were generated from the training partition only, by four methods run over 4 copies each.

Training method	Stored images
mild_gaussian_blur	80
mirror_and_crop	80
none	20
restrained_brightness	80
small_rotation	80

• restrained_brightness: brightness factor 0.75 to 0.90 or 1.10 to 1.30.
• small_rotation: 15 to 30 degrees either direction, corners filled with a representative scene colour.
• mirror_and_crop: horizontal mirror, then an 80 to 95 percent crop resized back to 224x224.
• mild_gaussian_blur: Gaussian radius 0.9 to 2.0 pixels at 224x224.

Label preservation drove the choice of methods. This label is defined by a bright emitting region against a darker surround, so transforms that manipulate tone can destroy it.

Contrast adjustment, used in the class notebook, is deliberately excluded: at a high factor it manufactures a bright-against-dark signal in an off photograph, and at a low factor it erases that signal from an on photograph. Brightness is retained but narrowed well inside the class notebook's range, which reaches 2.0 and can clip a dark scene's highlights into something resembling emission. A mirror-and-crop takes contrast's place because geometry cannot change whether a source is emitting.

Labels are copied from the parent image and never recomputed. The notebook asserts that every synthetic label matches its parent's.

Intended use and limitations

Suitable for teaching, augmentation practice and small-model prototyping. Not suitable for any real application.

• Validation and test hold a handful of images each. Accuracy on them moves in very large steps and indicates little.
• Thirty photographs from one person's surroundings, taken on one phone camera, do not represent light sources generally.
• Classes are uneven among the originals, with more on than off photographs, because several subjects were captured only while lit.
• Synthetic images carry no information beyond the training photographs. They add sampling density, not diversity, and a model evaluated on them would score optimistically.
• Padding to a square leaves neutral grey bars on every non-square frame. A model could in principle key on bar geometry, though the padding rule is identical across classes and splits.
• Subject grouping is inferred from capture order rather than recorded at collection time. The one-on-one-off check makes a wrong pairing unlikely but does not rule it out.

Ethical considerations

No people, faces or personally identifying information are included. All photographs are the author's own, taken of inanimate light fixtures and vehicle lights. The dataset carries no sensitive attributes and poses no foreseeable privacy risk.

License

CC BY 4.0.

AI usage disclosure

All thirty photographs were taken by the author. Subject selection, framing, the on/off labelling and the decision to exclude contrast augmentation were the author's.

An LLM assistant (Claude) was used to write and refine the notebook code, the galleries and this card, to inspect the collected files for class balance and count, and to propose the subject-pairing inference that the author verified. No image in this dataset was generated by a model. Synthetic images are produced by deterministic Pillow transforms in the accompanying notebook.