T0MYYY/nutrition5k-experiments

Nutrition5k — Reproduction & DPF-Nutrition checkpoints

PyTorch checkpoints (best.pt) from our Nutrition5k study, trained and evaluated under the official train/test splits.

• 🔬 Research repo (training, configs, full results): https://github.com/T0MYYY/Nutrition5k
• 📱 iOS deployment app: https://github.com/T0MYYY/CalBro
• 🤗 Deployment (Core ML) weights: https://huggingface.co/T0MYYY/dpf-nutrition

Checkpoints

Folder	Track	Backbone / init	Notes
dpf/best.pt	DPF-Nutrition	ImageNet init	Plain DPF-Nutrition (depth prediction + RGB-D fusion)
dpf_food2k/best.pt	DPF-Nutrition	Food2K init	DPF with Food2K-pretrained backbone
exp1/best.pt	Exp1 — portion-independent (per-gram), side-angle	InceptionV3	Paper Exp1 reproduction
exp1_convnext/best.pt	Exp1 ablation	ConvNeXt-Small (IN-22K→1K)	Backbone ablation vs InceptionV3
exp2/best.pt	Exp2 — direct prediction, side-angle	InceptionV3	Paper Exp2 reproduction
exp3/best.pt	Exp3 — direct, overhead RGB-D (depth channel)	InceptionV3	Mass estimation tracks/surpasses the paper
exp4/best.pt	Exp4 — MassRegressor + cal/g pipeline	InceptionV3	Overhead + volume scalar

See the research repo for per-experiment metrics (PMAE), configs, and the backbone-ablation analysis.

Citation

@inproceedings{thames2021nutrition5k,
  title = {Nutrition5k: Towards Automatic Nutritional Understanding of Generic Food},
  author = {Thames, Quin and Karpur, Arjun and Norris, Wade and Xia, Fangting and Panait, Liviu and Weyand, Tobias and Sim, Jack},
  booktitle = {CVPR}, year = {2021}
}
@article{han2023dpfnutrition,
  title = {DPF-Nutrition: Food Nutrition Estimation via Depth Prediction and Fusion},
  author = {Han, Yuzhe and Cheng, Qimin and Wu, Wenjin and Huang, Ziyang},
  journal = {Foods}, volume = {12}, number = {23}, pages = {4293}, year = {2023}, doi = {10.3390/foods12234293}
}