[NeurIPS 2023] Continuous Parametric Optical Flow

Abstract

In this paper, we present continuous parametric optical flow, a parametric representation of dense and continuous motion over arbitrary time interval. In contrast to existing discrete-time representations (i.e.flow in between consecutive frames), this new representation transforms the frame-to-frame pixel correspondences to dense continuous flow. In particular, we present a temporal-parametric model that employs B-splines to fit point trajectories using a limited number of frames. To further improve the stability and robustness of the trajectories, we also add an encoder with a neural ordinary differential equation (ODE) to represent features associated with specific times. We also contribute a synthetic dataset and introduce two evaluation perspectives to measure the accuracy and robustness of continuous flow estimation. Benefiting from the combination of explicit parametric modeling and implicit feature optimization, our model focuses on motion continuity and outperforms than the flow-based and point-tracking approaches for fitting long-term and variable sequences.

Requirements

Python 3.8.10 with basic conda environment. Install the requirements as follow:

pip install -r requirements.txt

Evaluation

1. Dataset Preparation

   • Download the TAP-Vid-DAVIS. Put the single *.pkl file into folder ./datasets/tap_vid_davis
   • Refer to the instruction to generate TAP-Vid-Kinetics.Put all *.pkl files into folder ./datasets/tap_vid_kinetics
   • Note: please use the dataset_mode to decide which one to evaluate (davis for Vid-TAP-DAVIS & kinetics for Vid-TAP-Kinetics)

2. Metric Selection

   • Accuracy(ade_all) & Smoothness(TRMSE) analysis using the ade_rmse
   • visibility(ade_vis) & occlusion(ade_occ) analysis using the ade_occ

3. Inference

   • Example: analyse the overall trajectory quality(accuracy & smoothness) on TAP-Vid-DAVIS

     python eval_real_scene.py --dataset_mode davis --method ade_rmse

4. Download the weight cp_flow_30.pth and put it into folder ./checkpoints

Training

We recommend parallel training with multiple GPUs. The dataset_dir and save_path denote the path where the training dataset is stored and the directory where the model weights are saved, respectively. A training sample is below:

• Example: use a single machine with 4 GPUs for training

  CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --node_rank=0 --master_addr="YOUR_IP" --master_port="YOUR_PORT"  kubric_sparse_train_all.py --dataset_dir "YOUR_DATASET_PATH" --save_path "YOUR_SAVE_PATH"

  We also provide the sampler code for variable-length training. Please refer to kubric_dataset.py.

Due to the large-scale transferring, we will upload the whole training dataset simulated by kubric later.

Acknowledgement

Thanks for the inspiration from the following work:

• Vid-ODE: Continuous-Time Video Generation with Neural Ordinary Differential Equation
• Raft: Recurrent all-pairs field transforms for optical flow
• Tap-vid: A benchmark for tracking any point in a video