From 17e157edc455c8a54ed20429cdaab89c0cadb9c9 Mon Sep 17 00:00:00 2001 From: Georg Hess Date: Tue, 2 Apr 2024 15:17:54 +0200 Subject: [PATCH] add missing closing tag (#20) --- _publications/closing-real2sim-gap/closing_real2sim_gap.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/_publications/closing-real2sim-gap/closing_real2sim_gap.md b/_publications/closing-real2sim-gap/closing_real2sim_gap.md index 1472380..c0e54fb 100644 --- a/_publications/closing-real2sim-gap/closing_real2sim_gap.md +++ b/_publications/closing-real2sim-gap/closing_real2sim_gap.md @@ -44,7 +44,7 @@ To reduce the real2sim gap, we propose to make perception models more robust tow # Results -We evaluate three 3D object detectors and an online mapping model on both real and NeRF data. We find that without fine-tuning, the models perform significantly worse on NeRF data compared to real data. However, after fine-tuning, all models improve notably on the NeRF data, and in some cases even increase in performance on real data. This indicates that the models have become more robust toward NeRF artifacts. We also find that the img2img method is the most effective augmentation method, followed by NeRF-data and image augmentation. +We evaluate three 3D object detectors and an online mapping model on both real and NeRF data. We find that without fine-tuning, the models perform significantly worse on NeRF data compared to real data. However, after fine-tuning, all models improve notably on the NeRF data, and in some cases even increase in performance on real data. This indicates that the models have become more robust toward NeRF artifacts. We also find that the img2img method is the most effective augmentation method, followed by NeRF-data and image augmentation. ## 3D Object Detection