giosanchez0208/clothing-color-recognition

Clothing color recognition

Predicts a probability distribution over 13 color categories for a garment, rather than assigning a single label. A shirt that is two-thirds navy and one-third white has a correct answer that no single label can express, so the target is a distribution and the loss is KL divergence, which is Label Distribution Learning in the sense of Geng (2016).

Trained on 28,860 procedurally generated images and zero real photographs.

Architecture	MobileNetV3-Small, 13-way head
Parameters	1.53 M
Size	6.0 MB FP32, 4.2 MB INT8, 5.85 MB ONNX
Teacher	ResNet-50, distilled at T=4.0, alpha=0.7
Classes	red, orange, yellow, green, blue, violet, purple, white, gray, black, pink, brown, olive

Try it

Browser demo, model runs client-side: https://huggingface.co/spaces/giosanchez0208/clothing-color-recognition

Results

Held-out test split, 2,000 images, read once, backgrounds drawn from a pool disjoint from train, validation, and the probe set.

Model	KL	top-1	MAE	Size	CPU
Teacher, ResNet-50	0.6226	63.2%	0.0499	90.1 MB	45.4 ms
Student FP32	0.4799	67.3%	0.0423	6.0 MB	6.2 ms
Student INT8	0.4800	67.4%	0.0423	4.2 MB	6.9 ms

The student beats its teacher on every metric, which V2 of this project also observed but measured with backgrounds shared between train and validation. Here the split is honest and the inversion holds: 23% lower KL and 4.2 points higher top-1 than the model it was distilled from. The likeliest mechanism is that MobileNetV3's depthwise-separable convolutions impose a stronger inductive bias than ResNet-50's deeper parameterization, which regularizes better on a 13-class problem where the teacher is mildly overfitting the synthetic distribution.

INT8 dynamic quantization is free here: KL moves by 0.0001 and top-1 goes up a tenth of a point.

Files

File	What it is
distilB_student.onnx	ONNX, opset 17, dynamic batch. Named I/O: image -> color_logits. Use this for the browser or cross-platform inference.
distilB_student_fp32.pth	PyTorch state dict, FP32.
distilB_student_int8.pth	PyTorch state dict, dynamically quantized. CPU only.

The checkpoints carry an architecture field so a loader does not have to guess which of the three it holds.

Input format, which is not a plain resize

The model was trained on a composite, not on photographs: a 112x112 crop of the garment region pasted into the middle of a 224x224 wider-context crop, then ImageNet-normalized. Feeding it a plain resized photo produces confident nonsense. The reference implementation is compose_input() in scripts/predict.py in the source repository, and a JavaScript port is in the Space.

IMG_SIZE, INNER = 224, 112
MEAN = [0.485, 0.456, 0.406]
STD  = [0.229, 0.224, 0.225]

Outputs are raw logits over the 13 classes in the order listed above. Apply softmax to get the distribution.

Limitations

No real-data validation. Every number above is measured on synthetic data. The central claim, that a model trained purely on procedurally generated images transfers to photographs, is currently untested at scale.

Weakest at category boundaries. Per-axis probing at convergence shows hue carrying 4.5 times the cost of the next axis. The known failure is pale green read as yellow: the model's learned green/yellow boundary sits near 135 degrees of hue while the color library puts it at 117.7, so a band that is labeled green in training gets classified yellow at inference. The cause is the generator sampling colors with inverse-Mahalanobis weights, which under-represents the region right against the boundary; only 2 of 25 library greens sit below 135 degrees, taking 4.1% of green draws. Gray against white fails similarly.

Color constancy is learned rather than preprocessed. Applying white balance at inference gives no improvement, which is positive evidence the augmentation pipeline pushed the correction into the weights, but it also means the model has only ever seen the illuminant range the generator simulated.

Licensing and provenance

Project code is MIT. Two things ride along and are worth stating plainly:

• The backbones initialize from ImageNet-pretrained torchvision weights, whose terms permit non-commercial research use. Anyone reusing this model commercially should confirm the current ImageNet terms cover their situation.
• The training backgrounds derive from MIT Indoor Scene Recognition (Quattoni and Torralba, CVPR 2009). No background images are redistributed here.

The color taxonomy derives from Paul Centore's sRGB centroids for the ISCC-NBS Colour System, which originates in NBS Special Publication 440, a US Government publication in the public domain.

Citation

@software{sanchez2026clothingcolor,
  author = {Sanchez, Gio},
  title  = {Clothing Color Recognition with a Synthetic Dataset},
  year   = {2026},
  url    = {https://github.com/giosanchez0208/Clothing-Color-Recognition-ML-With-Synthetic-Dataset}
}

References

• Geng, X. (2016). Label Distribution Learning. IEEE TKDE 28(7), 1734-1748.
• Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the Knowledge in a Neural Network. arXiv:1503.02531
• Müller, R., Kornblith, S., & Hinton, G. (2019). When Does Label Smoothing Help? NeurIPS. arXiv:1906.02629
• Quattoni, A. & Torralba, A. (2009). Recognizing Indoor Scenes. CVPR.