I3D Baseline for Isolated Sign Language Recognition (i3d-48f-50fps-224px)

This repository contains the weights for the I3D baseline model evaluated under the Signer-Independent (SI) split on the LSFB-isol benchmark, as presented in the paper:

Sign Language Recognition with Video Foundation Models: V-JEPA Transfer and Adaptation
Giovanne Mariano and Helio Pedrini (Institute of Computing, University of Campinas - UNICAMP)
Paper / Repository: GitHub Code & Artifacts

---

Model Description

• Backbone: Inflated 3D ConvNet (I3D) based on an Inception-v1 backbone.
• Initialization: Pre-trained on ImageNet, Kinetics-600, and Charades.
• Task: Isolated Sign Language Recognition (ISLR) on French Belgian Sign Language (LSFB-isol) across 700 classes.
• Resolution: 224x224 px spatial crop.
• Temporal Configuration: 48 context frames sampled at the native 50 FPS.
• Aggregation: Per-frame logits interpolated back to original temporal length and max-pooled over time. The interpolation is kept for exact reproduction of the original LSFB-isol release paper.

---

Performance Metrics (700-Class LSFB-isol)

Evaluated under the Signer-Independent (SI) 5-fold split with held-out signers:

---

Hyperparameters & Training Setup

The hyperparameters were selected via Optuna on a reduced 50-class subset:

global:
  seed: 73
  model_name: i3d
execution:
  batch_size: 32
  accumulate_grad_batches: 1
  num_workers: 4
dataset:
  data_root: /path/to/lsfb-isol
  train_split: si_train
  val_split: si_val
  num_classes: 700
pipeline:
  max_frames: 80
  num_frames: 48
  crop_size: 224
  target_fps: 50.0
model:
  pretrained: imagenet_kinetics_charades
  dropout: 0.23
optimizer:
  name: sgd
  lr: 0.032
  momentum: 0.97
  weight_decay: 3.3e-06
training:
  max_epochs: 50
  warmup_epochs: 5
  scheduler: cosine

---

Usage Example

To load the model weights in PyTorch:

import torch
from huggingface_hub import hf_hub_download

# Download weights from Hugging Face Hub
repo_id = "giolucasd/i3d-48f-50fps-224px"
weights_path = hf_hub_download(repo_id=repo_id, filename="best_weights.pt")

# Load state dict
state_dict = torch.load(weights_path, map_location="cpu", weights_only=True)

---

Citation

@inproceedings{mariano2026sign,
  title={Sign Language Recognition with Video Foundation Models: V-JEPA Transfer and Adaptation},
  author={Giovanne Mariano and H{\'e}lio Pedrini},
  booktitle={Workshop de Vis{\~a}o Computacional (WVC)},
  year={2026}
}