Dedelayed

Works

[CVPR 2026] Dedelayed: Deleting Remote Inference Delay via On-Device Correction [paper] [archival code snapshot]

Installation

git clone https://github.com/InterDigitalInc/dedelayed.git
cd dedelayed

# Using uv:
uv sync

# Using pip:
python -m venv .venv
source .venv/bin/activate
pip install -e .

Quick Start (Demo)

uv run python dedelayed/apps/dedelayed_v1/scripts/demo.py samples/input/source_video.mp4 --model_name=dedelayed_v1_efficientvitl1_mstransformer2d_bdd100k

Quick Start (Inference)

import torch
from dedelayed.zoo import get_model

device = "cuda"
model_name = "dedelayed_v1_efficientvitl1_mstransformer2d_bdd100k"  # CVPR 2026
model = get_model(model_name, pretrained=True).eval().to(device=device)

x_remote = torch.rand((1, 3, 4, 720, 1248), device=device)
x_local = torch.rand((1, 3, 480, 832), device=device)
past_ticks = torch.full((1,), 5.0, device=device)

with torch.inference_mode():
    out_remote = model.remote_model(
        x_remote,
        past_ticks=past_ticks,
        x_local_size=x_local.shape[-2:],
    )
    out_local = model.local_model(
        x_local,
        downlink_features=out_remote["downlink_features"],
    )
    pred_mask = out_local["seg_logits"].argmax(dim=1)

Training

Configure datasets:

Set meta["hp"]["dataset"] to dataset path.

Then, run training:

python scripts/train.py

Dataset format

The reference dataloader assumes each Hugging Face sample looks roughly like this:

For frames i in [0, 15]:

• original_{i} – uncompressed RGB images for evaluation.
• near_lossless_{i} – current-frame RGB images.
• label_{i} / label_hq_{i} – semantic masks stored as single-channel PIL images using Cityscapes IDs (ignore index 255).

License

This project is distributed under the BSD license included in LICENSE. The notice explicitly states that no patent rights are granted; review the file before redistributing or modifying the software.

Citation

@inproceedings{jacobellis2026dedelayed,
  title     = {Dedelayed: Deleting Remote Inference Delay via On-Device Correction},
  author    = {Jacobellis, Dan and Ulhaq, Mateen and Racap{\'e}, Fabien and Choi, Hyomin and Yadwadkar, Neeraja J.},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026},
  note      = {To appear}
}