Downloads · 30 days
0
segformer-b2-clothes-onnx
AbrahamPJ/segformer-b2-clothes-onnx
segformer-b2-clothes-onnx is a machine learning model from AbrahamPJ. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
mattmdjaga/segformerb2clothes (SegFormer-B2 fine-tuned on ATR, 27.4M params) quantized to int8 with onnxruntime.quantization.quantizedynamic(QUInt8) for on-device CPU inference.
Access
Public
Updated Sep 23, 2026
Repo size
28.6 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.onnx28.6 MB · 100%
At a glance
- License
- other
- Access
- Public
- Created
- Sep 23, 2026
- Updated
- Sep 23, 2026
- SHA
- 8460f9e5
Model card
From the Hugging Face model README
ATR human parser, int8 ONNX (for Nightmare Mobile)
mattmdjaga/segformer_b2_clothes
(SegFormer-B2 fine-tuned on ATR, 27.4M params) quantized to int8 with
onnxruntime.quantization.quantize_dynamic(QUInt8) for on-device CPU inference.
- input
pixel_values[1,3,512,512]float32 — image squashed to 512x512 (not letterboxed), rescaled 1/255, then ImageNet mean/std. - output
logits[1,18,128,128]float32 — argmax over the 18 ATR labels.
Labels: Background, Hat, Hair, Sunglasses, Upper-clothes, Skirt, Pants, Dress, Belt, Left-shoe, Right-shoe, Face, Left-leg, Right-leg, Left-arm, Right-arm, Bag, Scarf.
Measured, Snapdragon 8 Elite (S25 Ultra), ORT 1.29.0 CPU, 4 threads
| fp32 | this (int8) | |
|---|---|---|
| bytes | 110,039,290 | 28,635,315 |
| inference, warm | 886-999 ms | 510-616 ms |
Label agreement with fp32 is 97.8-99.9% of pixels, and the union masks (clothes, skin) are IoU 0.97. Quantization moves pixels between adjacent labels (Dress vs Upper-clothes, a limb vs Bag) but never empties a mask — so grouped targets are stable and individual garment names are not.
Weights and license are upstream's; this repo only re-exports them.
- License
- other
- Created
- Sep 23, 2026
- Updated
- Sep 23, 2026