Skip to content

Latest commit

 

History

History
71 lines (44 loc) · 3.77 KB

README.md

File metadata and controls

71 lines (44 loc) · 3.77 KB

Dynamic-Object-Removal-and-Inpainting

A Visual SLAM system improved with Dynamic object removal and inpainting the scene.

The project aims to segment dynamic objects in the video, remove them from the frames, and further inpaint these frames with background as we don’t want dynamic objects to be used for camera localization and to be included in the final reconstructed map of the scene.This project focuses on two major aspects, Dynamic Object Segmentation and Video Inpainting. The Project tries to improve the SLAM system with efficient and accurate learning based methods built upon the baseline DynaSLAM- https://github.com/BertaBescos/DynaSLAM.

The Project is divided into two phases:

  1. Dynamic Object Detection and Semantic Segmentation
  2. Video Inpainting

Dynamic Object Detection and Semantic Segmentation

To achieve the dynamic object detection on certain dynamic object classes such as Humans, Birds etc. we fine tune a pre trained Transformer model called Segmenter building upon the baseline code (Source Code).

Install the environment according to above source and place the segmenter folder in the directory where the below notebooks are run. Also, please refer to instructions inside the ipynb files for the config files need to be replaced with source configs before performing transfer learning

The Integrated Code Implementation done to achieve this task can be accessed through Jupyter Notebook.

Also, the Loss Curves and Precision Curves are inferred for different datasets when fine tuned the pre trained Model. The Source Notebook Can be found here

Training and Validation curves :

Training loss curve

Train_loss

Validation loss curve

Validation_loss

Validation Mean IoU Curve

Validation_loss

Object Detection Results:

The Segmentation Results Before and After Fine Tuning:

Segmentation Result Before Fine Tuning

Video Inpainting

To improve the Background reconstruction quality after dynamic object removal upon the exisiting implementation of DynaSLAM we use Inpainting Technique called ProPainter building upon the baseline code (Source Code).

To run inference using Propainter, sequences from TUM RGBD Dataset/custom Data require preprocessing(Resizing,sequencing images etc).The code used for data processing is provided here.

Video Inpainting Results Results:

The Video Inpainting Results of an Image Compared with DynaSLAM:

Segmentation Result Before Fine Tuning

The Video result inferred on custom dataset:

masked input Inpainted Output

Acknowledgements:

We would like to acknowledge and refer the works below for building on our baseline code

  1. DynaSLAM: Tracking, Mapping, and Inpainting in Dynamic Scenes by Berta Bescos and Jose M. Facil and Javier Civera and Jose Neira (Source)
  2. Segmenter: Transformer for Semantic Segmentation by Strudel, Robin and Garcia, Ricardo and Laptev, Ivan and Schmid, Cordelia (Source)
  3. ProPainter: Improving Propagation and Transformer for Video Inpainting by Shangchen Zhou and Chongyi Li and Kelvin C. K. Chan and Chen Change Loy (Source)