deblurring-tsrn-rtvd

TSRN-RTVD

This repository contains a matched deployment pair for real-time, trajectory-conditioned video deblurring:

• TSRNN_BT1, the recurrent one-frame-look-ahead deblurrer;
• FastShiftRegressor, which predicts previous-to-center and next-to-center global pixel shifts from three blurry frames.

Files

configs/tsrnn_bt1.yaml
configs/fast_shift_small.yaml
checkpoints/tsrnn_bt1.pth
checkpoints/fast_shift_small.pt

The two configurations use the same field of view, reference depths, projection grid, and pose convention. Do not mix checkpoints and configs from experiments with different projection semantics.

Usage

from tsrn_rtvd import TSRNRTVD

model = TSRNRTVD.from_pretrained(
    "YOUR_USERNAME/deblurring-tsrn-rtvd",
    device="cuda",
)

For installation, camera demo, folder inference, training details, dataset layout, and geometric limitations, see the source repository: https://github.com/msu-video-group/tsrn-rtvd

Limitations

FastShift predicts one global 2D vector per temporal direction. Such vectors cannot exactly represent parallax, perspective rotation, rolling shutter, dynamic objects, spatially varying blur, or changing scene depth. Input height and width for TSRNN_BT1 must be divisible by 8.