Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Fine-grained Large Object Segmentation 1st Place Solution to 3D AI Challenge 2020 -- Instance Segmentation Track

Zehui Chen, Qiaofei Li, Feng Zhao

TL;DR

The paper tackles large-object instance segmentation on the 3D-FUTURE dataset, where conventional COCO-focused methods underperform. It systematically compares HTC, SOLOv2, and PointRend, and demonstrates that PointRend's coarse-to-fine, point-based rendering yields substantially finer masks for large objects. Through targeted enhancements (more points, larger backbones, P6-FPN, higher input resolution) and a five-model ensemble with Linear-Interpolation weighting and soft-NMS, the approach achieves state-of-the-art results on both validation and test sets ($79.17$ val, $77.38$ test). The work highlights the practical effectiveness of fine-grained segmentation for large indoor objects and secures 1st place in the 3D AI Challenge Instance Segmentation Track.

Abstract

This technical report introduces our solutions of Team 'FineGrainedSeg' for Instance Segmentation track in 3D AI Challenge 2020. In order to handle extremely large objects in 3D-FUTURE, we adopt PointRend as our basic framework, which outputs more fine-grained masks compared to HTC and SOLOv2. Our final submission is an ensemble of 5 PointRend models, which achieves the 1st place on both validation and test leaderboards. The code is available at https://github.com/zehuichen123/3DFuture_ins_seg.

Towards Fine-grained Large Object Segmentation 1st Place Solution to 3D AI Challenge 2020 -- Instance Segmentation Track

TL;DR

The paper tackles large-object instance segmentation on the 3D-FUTURE dataset, where conventional COCO-focused methods underperform. It systematically compares HTC, SOLOv2, and PointRend, and demonstrates that PointRend's coarse-to-fine, point-based rendering yields substantially finer masks for large objects. Through targeted enhancements (more points, larger backbones, P6-FPN, higher input resolution) and a five-model ensemble with Linear-Interpolation weighting and soft-NMS, the approach achieves state-of-the-art results on both validation and test sets ( val, test). The work highlights the practical effectiveness of fine-grained segmentation for large indoor objects and secures 1st place in the 3D AI Challenge Instance Segmentation Track.

Abstract

This technical report introduces our solutions of Team 'FineGrainedSeg' for Instance Segmentation track in 3D AI Challenge 2020. In order to handle extremely large objects in 3D-FUTURE, we adopt PointRend as our basic framework, which outputs more fine-grained masks compared to HTC and SOLOv2. Our final submission is an ensemble of 5 PointRend models, which achieves the 1st place on both validation and test leaderboards. The code is available at https://github.com/zehuichen123/3DFuture_ins_seg.

Paper Structure

This paper contains 12 sections, 1 equation, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Datasets
  3. Methods
  4. Hybrid Task Cascade
  5. SOLOv2
  6. PointRend
  7. Final Submission
  8. Implementation Details
  9. Test Performance
  10. Model Ensemble
  11. Visualization
  12. Conclusion

Figures (2)

  • Figure 1: Size distribution of bounding boxes in 3D-FUTURE and COCO. We randomly select 10,000 images for fair comparison. $x$ axis denotes the sqrt area of a bounding box and $y$ axis denotes the number of boxes within each corresponding interval.
  • Figure 2: Example of segmentation results on validation dataset from three best single models: (a)(d) HTC, (b)(e) SOLOv2 and (c)(f) PointRend. PointRend predicts masks with substantially finer details around object boundaries. All figures are best viewed digitally with zoom.