This report presents the winning solution for the Autonomous Driving Challenge 2023 - Online HD-map Construction. The solution, called MachMap, focuses on the vectorization pipeline and introduces an effective architecture that formulates HD-map construction as a point detection task in the bird-eye-view space. The authors first introduce a novel map-compaction scheme into their framework, which significantly reduces the number of vectorized points by 93% without compromising performance. They then propose a strong baseline approach that integrates a powerful CNN-based backbone called InternImage, a temporal-based instance decoder, and a well-designed point-mask coupling head. Additionally, they utilize an optional ensemble stage to further refine model predictions for improved performance. The authors evaluate their MachMap-tiny model with IN-1K initialization on the Argoverse2 benchmark and achieve a mean Average Precision (mAP) of 79.1. They further improve their model with MachMap-huge, which achieves the best mAP of 83.5 on the benchmark. Notably, MachMap outperforms all other online HD-map construction approaches on the final leaderboard by a significant margin of at least 9.8 mAP. Overall, this report provides detailed insights into the MachMap solution and its effectiveness in constructing compact HD maps for autonomous driving applications with superior performance compared to existing approaches.
Summary- There is a solution called MachMap that helps make maps for self-driving cars.
- It uses a special way to make the maps and makes them smaller without losing quality.
- The solution has different parts that work together to make the maps better.
- They tested the solution and it did really well compared to other ways of making maps.
- MachMap is the best way to make online HD-maps for self-driving cars.
Definitions
- Autonomous Driving Challenge 2023: A competition where people try to find the best solutions for self-driving cars.
- HD-map: A detailed map that self-driving cars use to know where they are and where they should go.
- Vectorization pipeline: A process of turning information into lines and shapes on a map.
- Bird-eye-view space: Looking at something from above, like a bird flying in the sky.
- Baseline approach: A starting point or basic method that can be improved upon.
- Ensemble stage: Adding more things together to make something even better.
MachMap: An Effective Solution for Autonomous Driving Challenge 2023 - Online HD-map Construction
Autonomous driving technology has been rapidly advancing in recent years, and the development of high-definition (HD) maps is an essential component of this progress. To further promote research in this area, the Autonomous Driving Challenge 2023 (ADC2023) was launched with a focus on online HD-map construction. The winning solution for ADC2023 was MachMap, which formulates HD-map construction as a point detection task in the bird-eye-view space and introduces an effective architecture to achieve superior performance compared to existing approaches. This article provides detailed insights into MachMap's vectorization pipeline and its effectiveness in constructing compact HD maps for autonomous driving applications.
Vectorization Pipeline
The authors first introduce a novel map-compaction scheme into their framework, which significantly reduces the number of vectorized points by 93% without compromising performance. This scheme consists of two components: InternImage and PointMask Coupling Head. InternImage is a powerful CNN backbone that utilizes temporal information from multiple frames to detect objects such as cars or pedestrians accurately and efficiently. The PointMask Coupling Head then takes these detected objects as input and produces a binary mask indicating whether each pixel should be included in the vectorized output or not. This masking step allows for more efficient data representation while preserving accuracy during inference time.
Additionally, they utilize an optional ensemble stage to further refine model predictions for improved performance. This ensemble approach combines multiple models trained on different datasets with different parameters to produce more accurate results than any single model could achieve alone.
Evaluation Results
The authors evaluate their MachMap-tiny model with IN-1K initialization on the Argoverse2 benchmark and achieve a mean Average Precision (mAP) of 79.1%. They further improve their model with MachMap-huge, which achieves the best mAP of 83.5 on the benchmark - outperforming all other online HD-map construction approaches by at least 9
This report introduces the 1st place winning solution for the Autonomous Driving Challenge 2023 - Online HD-map Construction. By delving into the vectorization pipeline, we elaborate an effective architecture, termed as MachMap, which formulates the task of HD-map construction as the point detection paradigm in the bird-eye-view space with an end-to-end manner. Firstly, we introduce a novel map-compaction scheme into our framework, leading to reducing the number of vectorized points by 93% without any expression performance degradation. Build upon the above process, we then follow the general query-based paradigm and propose a strong baseline with integrating a powerful CNN-based backbone like InternImage, a temporal-based instance decoder and a well-designed point-mask coupling head. Additionally, an extra optional ensemble stage is utilized to refine model predictions for better performance. Our MachMap-tiny with IN-1K initialization achieves a mAP of 79.1 on the Argoverse2 benchmark and the further improved MachMap-huge reaches the best mAP of 83.5, outperforming all the other online HD-map construction approaches on the final leaderboard with a distinct performance margin (> 9.8 mAP at least).
