--- title: Carlafox: Towards reliable open-source 3D perception description: Collabora presents Carlafox, a web-based CARLA visualizer used to automatically generate error-free synthetic datasets for 3D object detection. The post... canonical: https://daily.dev/posts/carlafox-towards-reliable-open-source-3d-perception-trxxeqgrk twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Carlafox: Towards reliable open-source 3D perception | daily.dev og:description: Collabora presents Carlafox, a web-based CARLA visualizer used to automatically generate error-free synthetic datasets for 3D object detection. The post... og:url: https://daily.dev/posts/carlafox-towards-reliable-open-source-3d-perception-trxxeqgrk og:image: https://api.daily.dev/og/posts/tRxxeQgRk.png og:image:alt: Carlafox: Towards reliable open-source 3D perception og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Carlafox: Towards reliable open-source 3D perception **[Collabora Blog](https://daily.dev/sources/collabora)** · 6 min read · 0 upvotes · 0 comments ## Summary Collabora presents Carlafox, a web-based CARLA visualizer used to automatically generate error-free synthetic datasets for 3D object detection. The post explains how labeling errors in real-world datasets degrade detector accuracy, and how synthetic data from CARLA (in KITTI format) can reduce reliance on expensive real-world data. The generated dataset contains 12,807 cars, 10,252 pedestrians, and 11,624 cyclists with 2D/3D bounding box annotations and LiDAR+camera sensor data. Several LiDAR-based models (PointRCNN, PVRCNN, SFA3D, MVXNet) were trained on this dataset, with SFA3D selected for fine-tuning due to its speed and memory efficiency. The SFA3D PyTorch model was then converted to ONNX and optimized with TensorRT (FP16 quantization) for inference, with benchmarks showing TensorRT outperforms TVM and plain ONNX on RTX 2080Ti. The model was integrated into Carlafox for side-by-side comparison with CARLA ground truth. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is the dataset size generated by Carlafox using CARLA for 3D object detection? The Carlafox-generated CARLA synthetic dataset contains 12,807 cars, 10,252 pedestrians, and 11,624 cyclists. It includes 2D and 3D bounding box annotations for Car, Pedestrian, and Cyclist classes, along with LiDAR and camera sensor data in KITTI format, plus sensor calibration matrices. The LiDAR and camera configurations match those used to generate the original KITTI dataset. _Teams building 3D perception pipelines track synthetic dataset approaches like this on daily.dev._ ### Why was SFA3D chosen over PointRCNN, PVRCNN, and MVXNet for fine-tuning on the CARLA synthetic dataset? SFA3D was selected for fine-tuning because it is faster and uses less memory without significant loss in performance compared to PointRCNN, PVRCNN, and MVXNet. The other models were only trained to baseline and not further optimized; any of them could potentially outperform SFA3D with additional tuning. SFA3D uses a ResNet-based Keypoint Feature Pyramid Network operating on bird's-eye-view LiDAR maps. _Developers choosing between LiDAR-based 3D detectors for embedded or real-time use find comparisons like this on daily.dev._ ### How do I convert a SFA3D PyTorch model to TensorRT for inference optimization? Convert the SFA3D PyTorch model to ONNX first, then use the TensorRT ONNX parser to produce a TensorRT engine. Alternatively, you can write the network directly in TensorRT's network-definition API, bypassing ONNX, but this is time-intensive with negligible speed benefit on desktop GPUs — though it may be more efficient on embedded devices like Jetson Nano. FP16 quantization further boosts throughput on RTX 2080Ti. _Engineers optimizing 3D perception inference for embedded hardware follow deployment trade-offs like these on daily.dev._ ## Similar posts on daily.dev - [One API to Rule Them All: py123d \+ FiftyOne for Autonomous Driving Data](https://daily.dev/posts/one-api-to-rule-them-all-py123d-fiftyone-for-autonomous-driving-data-it8xwsrm5) · Medium · 0 upvotes · 0 comments - [Into the Omniverse: Three Workflows for Improving Vision AI Agent Accuracy With Synthetic Data and Fine-Tuning](https://daily.dev/posts/into-the-omniverse-three-workflows-for-improving-vision-ai-agent-accuracy-with-synthetic-data-and-f-4egob4ssz) · NVIDIA · 0 upvotes · 0 comments - [Scale AV Perception Across Vehicle Platforms with NVIDIA Omniverse NuRec](https://daily.dev/posts/scale-av-perception-across-vehicle-platforms-with-nvidia-omniverse-nurec-vfuubxqw6) · NVIDIA Developer · 0 upvotes · 0 comments - [Falcon Perception](https://daily.dev/posts/falcon-perception-x4ei6gikw) · Hugging Face · 0 upvotes · 0 comments - [How Does AI Learn to See in 3D and Understand Space?](https://daily.dev/posts/how-does-ai-learn-to-see-in-3d-and-understand-space--cmvmufvtv) · Towards Data Science · 0 upvotes · 0 comments --- Tags: [#tech-news](https://daily.dev/tags/tech-news) [View this post on daily.dev](https://daily.dev/posts/carlafox-towards-reliable-open-source-3d-perception-trxxeqgrk) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Carlafox: Towards reliable open-source 3D perception","url":"https://daily.dev/posts/carlafox-towards-reliable-open-source-3d-perception-trxxeqgrk","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/carlafox-towards-reliable-open-source-3d-perception-trxxeqgrk"},"datePublished":"2026-08-11T08:55:49.165Z","dateModified":"2026-08-11T09:12:16.631Z","description":"Collabora presents Carlafox, a web-based CARLA visualizer used to automatically generate error-free synthetic datasets for 3D object detection. The post...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/2a736220749d2cede790e51f7f66e435?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/2a736220749d2cede790e51f7f66e435?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collabora Blog","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collabora Blog","logo":"https://media.daily.dev/image/upload/s--ZGZ0FjqS--/f_auto,q_auto/v1786438306/logos/collabora?_a=BAMAMicg0","url":"https://daily.dev/sources/collabora"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/carlafox-towards-reliable-open-source-3d-perception-trxxeqgrk","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"tech-news","timeRequired":"PT6M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collabora Blog","item":"https://daily.dev/sources/collabora"},{"@type":"ListItem","position":3,"name":"Carlafox: Towards reliable open-source 3D perception"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/carlafox-towards-reliable-open-source-3d-perception-trxxeqgrk#faq","mainEntity":[{"@type":"Question","name":"What is the dataset size generated by Carlafox using CARLA for 3D object detection?","acceptedAnswer":{"@type":"Answer","text":"The Carlafox-generated CARLA synthetic dataset contains 12,807 cars, 10,252 pedestrians, and 11,624 cyclists. It includes 2D and 3D bounding box annotations for Car, Pedestrian, and Cyclist classes, along with LiDAR and camera sensor data in KITTI format, plus sensor calibration matrices. The LiDAR and camera configurations match those used to generate the original KITTI dataset. Teams building 3D perception pipelines track synthetic dataset approaches like this on daily.dev."}},{"@type":"Question","name":"Why was SFA3D chosen over PointRCNN, PVRCNN, and MVXNet for fine-tuning on the CARLA synthetic dataset?","acceptedAnswer":{"@type":"Answer","text":"SFA3D was selected for fine-tuning because it is faster and uses less memory without significant loss in performance compared to PointRCNN, PVRCNN, and MVXNet. The other models were only trained to baseline and not further optimized; any of them could potentially outperform SFA3D with additional tuning. SFA3D uses a ResNet-based Keypoint Feature Pyramid Network operating on bird's-eye-view LiDAR maps. Developers choosing between LiDAR-based 3D detectors for embedded or real-time use find comparisons like this on daily.dev."}},{"@type":"Question","name":"How do I convert a SFA3D PyTorch model to TensorRT for inference optimization?","acceptedAnswer":{"@type":"Answer","text":"Convert the SFA3D PyTorch model to ONNX first, then use the TensorRT ONNX parser to produce a TensorRT engine. Alternatively, you can write the network directly in TensorRT's network-definition API, bypassing ONNX, but this is time-intensive with negligible speed benefit on desktop GPUs — though it may be more efficient on embedded devices like Jetson Nano. FP16 quantization further boosts throughput on RTX 2080Ti. Engineers optimizing 3D perception inference for embedded hardware follow deployment trade-offs like these on daily.dev."}}]} ```