PUC

Phone Usage Classifier (PUC) is a three-class image classification pipeline for understanding how people interact with smartphones.

• classid=0 (no_action): No interaction with a smartphone.

• classid=1 (point_somewhere): Pointing the smartphone somewhere other than the camera.

• classid=2 (point): Pointing the smartphone towards the camera.

  output_.mp4

• Advanced Applications - Estimating whether or not the user is looking at their smartphone

  output__.mp4

  Variant	Size	F1	CPU inference latency	ONNX
  P	115 KB	0.9160	0.24 ms	Download
  N	176 KB	0.9337	0.39 ms	Download
  T	280 KB	0.9468	0.51 ms	Download
  S	495 KB	0.9672	0.66 ms	Download
  C	876 KB	0.9722	0.73 ms	Download
  M	1.7 MB	0.9774	0.86 ms	Download
  L	6.4 MB	0.9944	1.07 ms	Download

no action	no action	point somewhere	point somewhere	point	point

git clone https://github.com/PINTO0309/PUC.git && cd PUC
curl -LsSf https://astral.sh/uv/install.sh | sh
uv sync
source .venv/bin/activate

uv run python demo_puc.py \
-v 0 \
-pm puc_l_48x48.onnx \
-dlr -dnm -dgm -dhm \
-ep cuda

uv run python demo_puc.py \
-v 0 \
-pm puc_l_48x48.onnx \
-dlr -dnm -dgm -dhm \
-ep tensorrt

uv run python 01_data_prep_realdata.py

uv run python 01_data_prep_realdata.py \
--input-image-dir real_images \
--start-folder 1001 \
--allow-multi-body

uv run python 02_make_parquet.py --overwrite

Split counts:
  train: 27174
    val: 3021
Label counts:
         no_action: 12425
   point_somewhere: 8885
             point: 8885

• Use the labeled image folders under data/no_action, data/point_somewhere, and data/point.

• 02_make_parquet.py writes pre-defined train/val splits into data/dataset.parquet using an image-level 9:1 split per class.

• The training loop relies on BCEWithLogitsLoss plus class-balanced pos_weight to stabilise optimisation under class imbalance; inference produces sigmoid probabilities. Use --train_resampling weighted to switch on the previous WeightedRandomSampler behaviour, or --train_resampling balanced to physically duplicate minority classes before shuffling.

• Training history, validation metrics, optional test predictions, checkpoints, configuration JSON, and ONNX exports are produced automatically.

• Per-epoch checkpoints named like puc_epoch_0001.pt are retained (latest 10), as well as the best checkpoints named puc_best_epoch0004_f1_0.9321.pt (also latest 10).

• The backbone can be switched with --arch_variant. Supported combinations with --head_variant are:

  --arch_variant	Default (--head_variant auto)	Explicitly selectable heads	Remarks
  baseline	avg	avg, avgmax_mlp	When using transformer/mlp_mixer, you need to adjust the height and width of the feature map so that they are divisible by --token_mixer_grid (if left as is, an exception will occur during ONNX conversion or inference).
  inverted_se	avgmax_mlp	avg, avgmax_mlp	When using transformer/mlp_mixer, it is necessary to adjust --token_mixer_grid as above.
  convnext	transformer	avg, avgmax_mlp, transformer, mlp_mixer	For token mixer heads, the feature map dimensions must be divisible by --token_mixer_grid (default 2x3).

• The classification head is selected with --head_variant (avg, avgmax_mlp, transformer, mlp_mixer, or auto which derives a sensible default from the backbone).

• Pass --rgb_to_yuv_to_y to convert RGB crops to YUV, keep only the Y (luma) channel inside the network, and train a single-channel stem without modifying the dataloader.

• Alternatively, use --rgb_to_lab or --rgb_to_luv to convert inputs to CIE Lab/Luv (3-channel) before the stem; these options are mutually exclusive with each other and with --rgb_to_yuv_to_y.

• Mixed precision can be enabled with --use_amp when CUDA is available.

• Resume training with --resume path/to/puc_epoch_XXXX.pt; all optimiser/scheduler/AMP states and history are restored.

• Loss/accuracy/F1 metrics are logged to TensorBoard under output_dir, and tqdm progress bars expose per-epoch progress for train/val/test loops.

Baseline depthwise-separable CNN:

SIZE=48x48
uv run python -m puc train \
--data_root data/dataset.parquet \
--output_dir runs/puc_${SIZE} \
--epochs 100 \
--batch_size 256 \
--train_resampling balanced \
--image_size ${SIZE} \
--base_channels 32 \
--num_blocks 4 \
--arch_variant baseline \
--seed 42 \
--device auto \
--use_amp

Inverted residual + SE variant (recommended for higher capacity):

SIZE=48x48
VAR=s
uv run python -m puc train \
--data_root data/dataset.parquet \
--output_dir runs/puc_is_${VAR}_${SIZE} \
--epochs 100 \
--batch_size 256 \
--train_resampling balanced \
--image_size ${SIZE} \
--base_channels 32 \
--num_blocks 4 \
--arch_variant inverted_se \
--head_variant avgmax_mlp \
--seed 42 \
--device auto \
--use_amp

ConvNeXt-style backbone with transformer head over pooled tokens:

SIZE=48x48
uv run python -m puc train \
--data_root data/dataset.parquet \
--output_dir runs/puc_convnext_${SIZE} \
--epochs 100 \
--batch_size 256 \
--train_resampling balanced \
--image_size ${SIZE} \
--base_channels 32 \
--num_blocks 4 \
--arch_variant convnext \
--head_variant transformer \
--token_mixer_grid 2x2 \
--seed 42 \
--device auto \
--use_amp

• Outputs include the latest 10 puc_epoch_*.pt, the latest 10 puc_best_epochXXXX_f1_YYYY.pt (highest validation F1, or training F1 when no validation split), history.json, summary.json, optional test_predictions.csv, and train.log.
• After every epoch a confusion matrix and ROC curve are saved under runs/puc/diagnostics/<split>/confusion_<split>_epochXXXX.png and roc_<split>_epochXXXX.png.
• --image_size accepts either a single integer for square crops (e.g. --image_size 48) or HEIGHTxWIDTH to resize non-square frames (e.g. --image_size 64x48).
• Add --resume <checkpoint> to continue from an earlier epoch. Remember that --epochs indicates the desired total epoch count (e.g. resuming --epochs 40 after training to epoch 30 will run 10 additional epochs).
• Launch TensorBoard with:

  tensorboard --logdir runs/puc

uv run python -m puc exportonnx \
--checkpoint runs/puc_is_s_48x48/puc_best_epoch0049_f1_0.9939.pt \
--output puc_s_48x48.onnx \
--opset 17

• The saved graph exposes images as input and prob_pointing as output (batch dimension is dynamic); probabilities can be consumed directly.
• After exporting, the tool runs onnxsim for simplification and rewrites any remaining BatchNormalization nodes into affine Mul/Add primitives. If simplification fails, a warning is emitted and the unsimplified model is preserved.

1. VSDLM: Visual-only speech detection driven by lip movements - MIT License
2. OCEC: Open closed eyes classification. Ultra-fast wink and blink estimation model - MIT License
3. PGC: Ultrafast pointing gesture classification - MIT License
4. SC: Ultrafast sitting classification - MIT License
5. PUC: Phone Usage Classifier is a three-class image classification pipeline for understanding how people interact with smartphones - MIT License
6. HSC: Happy smile classifier - MIT License
7. WHC: Waving Hand Classification - MIT License
8. UHD: Ultra-lightweight human detection - MIT License
9. MWC: Mask wearing classifier - MIT License
10. SGC: Classification of wearing vs. not wearing sunglasses. 48x48. - MIT License
11. HHC: Head Hat Classification. HHC is a binary classifier for cropped head images. 48x48. - MIT License
12. BPC: Background Plain classification. 48x48. - MIT License
13. PPC: Binary classification to determine whether the subject is holding a smartphone. 48x48 RGB image. - MIT License

If you find this project useful, please consider citing:

@software{hyodo2025puc,
  author    = {Katsuya Hyodo},
  title     = {PINTO0309/PUC},
  month     = {11},
  year      = {2025},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.17666420},
  url       = {https://github.com/PINTO0309/puc},
  abstract  = {Phone Usage Classifier (PUC) is a three-class image classification pipeline for understanding how people
interact with smartphones.},
}

• https://github.com/PINTO0309/PINTO_model_zoo/tree/main/472_DEIMv2-Wholebody34: Apache 2.0 License

  @software{DEIMv2-Wholebody34,
    author={Katsuya Hyodo},
    title={Lightweight human detection models generated on high-quality human data sets. It can detect objects with high accuracy and speed in a total of 28 classes: body, adult, child, male, female, body_with_wheelchair, body_with_crutches, head, front, right-front, right-side, right-back, back, left-back, left-side, left-front, face, eye, nose, mouth, ear, collarbone, shoulder, solar_plexus, elbow, wrist, hand, hand_left, hand_right, abdomen, hip_joint, knee, ankle, foot.},
    url={https://github.com/PINTO0309/PINTO_model_zoo/tree/main/472_DEIMv2-Wholebody34},
    year={2025},
    month={10},
    doi={10.5281/zenodo.17625710}
  }

• https://github.com/PINTO0309/bbalg: MIT License