CenterDisks: Real-time instance segmentation with disk covering

TL;DR

CenterDisks tackles real-time instance segmentation by representing object masks with a fixed set of disks whose centers and radii are learned without direct supervision on disk parameters. Masks are produced as sums of 2D Gaussians, enabling end-to-end training with standard loss functions and a CenterNet-inspired architecture that predicts object centers, offsets, depths, and disk parameters. The method achieves competitive real-time performance on urban driving datasets (Cityscapes, IDD, KITTI) while maintaining low parameter counts, and ablation studies identify 16 disks with unique radii and a Dice loss as an effective configuration. This disk-covering approach offers a lightweight alternative to pixel-perfect masks, balancing speed and accuracy for autonomous perception and similar applications. Future work could extend the representation to ellipses or more complex shapes to better capture object contours and holes, widening applicability beyond the current datasets.

Abstract

Increasing the accuracy of instance segmentation methods is often done at the expense of speed. Using coarser representations, we can reduce the number of parameters and thus obtain real-time masks. In this paper, we take inspiration from the set cover problem to predict mask approximations. Given ground-truth binary masks of objects of interest as training input, our method learns to predict the approximate coverage of these objects by disks without supervision on their location or radius. Each object is represented by a fixed number of disks with different radii. In the learning phase, we consider the radius as proportional to a standard deviation in order to compute the error to propagate on a set of two-dimensional Gaussian functions rather than disks. We trained and tested our instance segmentation method on challenging datasets showing dense urban settings with various road users. Our method achieve state-of-the art results on the IDD and KITTI dataset with an inference time of 0.040 s on a single RTX 3090 GPU.

Figures (6)

• Figure 1: Disk covering on a Cityscapes image crop. The set of covering disks is composed of 24 disks with identical radius.
• Figure 2: Architecture for CenterDisks. The backbone is represented here as an Houglass backbone newell_stacked_2016. The five heads predict the heatmap for object centers, the offsets from this center, the relative depths, and the parameters of the disk sets. The number of parameters displayed on this figure is only as an illustration. For implementation details, please refer to section \ref{['sec:implementation']} and to the code provided.
• Figure 3: Thresholding with $\alpha = 0.5$ on the predicted set of disks at inference phase.
• Figure 4: Results on the Cityscapes validation set by categories.
• Figure 5: Qualitative results on the KITTI test set. Comparison with state-of-the-art instance segmentation methods. From top to bottom: CenterDisks (our method), Mask-RCNN he_mask_2017, Segment Anything kirillov_segment_2023, SparseInst cheng_sparse_2022. We used the pre-trained models provided by the authors, without fine-tuning them on KITTI. Best viewed on a screen.