TimeLens-XL: Real-time Event-based Video Frame Interpolation with Large Motion

[Project Page] [Paper Link] [Dataset]

This is the codebase for our paper TimeLens-XL, which has been accepted by ECCV2024. In this paper, we propose a lightweight netowrk for Event-VFI which is able to interpoalte for large motions in realtime at 720p resolution. We provide pretrained weights, code for evaluation for both our TLXNet, TLXNet+, and state-of-the-art methods.

Quick Start

Environment

Please download dependency packages by

pip install -r requirements.txt

Dataset

Download supported datasets and change path files in params/Paths accordingly. Currently we support:

1. Adobe dataset;
2. GOPRO dataset;
3. BSERGB dataset;
4. HQEVFI dataset [Google Drive];

Please notice that 1 and 2 are synthetic datasets. We first interpolate Adobe and GOPRO dataset to 1920 FPS (x8) with RIFE, and then synthesis events based on v2e.

Compared with original v2e, we made several changes on the simulator:

1. We remove image resize to make the RGB and event stream the same size, as we do not focus only on davis camera with fixed resolution;
2. We removed the random disturbance on time as we already have temporal dense RGB frames.

Please re-generate the synthetic dataset according to v2e simulator or we will release our simulator and synthetic datasets shortly.

Pretrained weights

on HQ-EVFI dataset: TLXNet+

on BSERGB dataset: TLXNet+ x4 interpolation

Evaluation on HQ-EVFI

python run_network.py --param_name traintest_RC_smallmix_lpips --model_name Expv8_large --model_pretrained ./weights/Expv8_large_HQEVFI.pt --skip_training

Training on HQ-EVFI

Just remove --skip_training of evaluation code.