Research map / Detection and segmentation / CNN features
GauCho: Gaussian Distributions with Cholesky Decomposition for Oriented Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GauCho, a regression head that predicts Cholesky parameters of a 2D Gaussian for oriented object detection, plus oriented ellipses as an output representation, and evaluates it on DOTA, HRSC and UCAS-AOD with several detectors.
Problem
Oriented object detection with OBB regression heads suffers from angular boundary discontinuity and encoding ambiguity for circular objects, and Gaussian-based losses still inherit these issues at inference.
Approach
Regress the lower-triangular Cholesky factor (alpha, beta, gamma) of the target Gaussian covariance directly from the network, a bijective and continuous mapping to Gaussians that is compatible with any Gaussian-based loss, and decode oriented ellipses instead of OBBs.
Key results
- On DOTA v1.0 single-scale, FCOS with GauCho-ProbIoU reaches 40.65 AP vs 39.80 for the OBB head with ProbIoU.
- On DOTA v1.5 with FCOS, GauCho improves AP50 by about 1.1% on average across GWD, KLD and ProbIoU losses.
- With multiscale testing, FCOS-GauCho reaches 78.85 AP50 on DOTA v1.0, slightly above CenterNet-ACM (78.53).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the OBB regression head with GauCho using ProbIoU | FCOS R-50 · HRSC | AP: 87.27 -> 87.58 | Helped |
| Replace the OBB regression head with GauCho using ProbIoU | FCOS R-50 · UCAS-AOD (OBB) | AP: 52.27 -> 53.03 | Helped |
| Replace the OBB regression head with GauCho using ProbIoU | FCOS R-50 · UCAS-AOD (OE) | AP: 59.27 -> 59.08 | Hurt |
| Replace the OBB regression head with GauCho using ProbIoU | FCOS R-50 · DOTA v1.0 | AP: 39.80 -> 40.65 | Helped |
| Replace the OBB regression head with GauCho using GWD | FCOS R-50 · DOTA v1.5 | AP50: 62.65 -> 64.00 (+1.35) | Helped |
| Replace the OBB regression head with GauCho using KLD | FCOS R-50 · DOTA v1.5 | AP50: 64.49 -> 65.51 (+1.02) | Helped |
| Replace the OBB regression head with GauCho using ProbIoU | FCOS R-50 · DOTA v1.5 | AP50: 64.26 -> 65.09 (+0.83) | Helped |
| Use GauCho instead of an OBB head | FCOS · DOTA v1.0 (multiscale training/testing) | AP50 | Helped |
| Use GauCho with RoI Transformer | RoI Transformer · DOTA v1.0 (multiscale training/testing) | AP50: 79.45 -> 80.61 | Helped |
| Use GauCho with FCOS ResNet101 | FCOS · DOTA v1.5 (multiscale training/testing) | mAP: 71.99 -> 73.56 | Helped |
| Represent final detections as OEs instead of OBBs | FCOS on UCAS-AOD · UCAS-AOD | AP75: 55.35/65.27 | Helped |
| Replace the OBB head with GauCho | FCOS on HRSC · HRSC | orientation error: 1.36° -> 1.11° | Helped |
Surprising findings
- On HRSC and UCAS-AOD, GauCho and OBB heads perform similarly, and OBB-based results are mostly comparable to GauCho for most detectors on DOTA v1.0 apart from FCOS.
Limitations
- Still suffers from decoding ambiguity for square-like objects, where orientation cannot be recovered from the Gaussian.
- Hyperparameters were inherited from OBB-head default configs and not tuned for GauCho.
- Gains over OBB heads are often small or within run-to-run variation except for FCOS on DOTA.
Experimental setup
Datasets: DOTA v1.0 DOTA v1.5 HRSC 2016 UCAS-AOD
Benchmarks: DOTA v1.0 DOTA v1.5 HRSC 2016 UCAS-AOD
Metrics: AP50 AP75 AP mAP orientation error
Baselines: OBB head with GWD OBB head with KLD OBB head with ProbIoU CenterNet-ACM RoI-Transformer-ACM DAFNe PSCD KFIoU DCL CSL
Builds on: FCOS RetinaNet R3Det RoI-Transformer ProbIoU
Closest papers by meaning
- Rethinking Boundary Discontinuity Problem for Oriented Object Detection (CVPR 2024 · similarity 0.85)
- ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects (CVPR 2025 · similarity 0.85)
- ReDiffDet: Rotation-equivariant Diffusion Model for Oriented Object Detection (CVPR 2025 · similarity 0.85)
- Theoretically Achieving Continuous Representation of Oriented Bounding Boxes (CVPR 2024 · similarity 0.85)
- FRED: Towards a Full Rotation-Equivariance in Aerial Image Object Detection (AAAI 2024 · similarity 0.84)
- Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.