June 2022
tl;dr: Early camera-lidar fusion in BEV space.
BEVFusion breaks the long-lasting common practice that point-level fusion is the golden choice for multi-sensor perception systems.
BEVFusion's main contribution seems to be the efficient implementation of the Voxel Pooling operation. The speed bottleneck in voxelpooling is also noted and solved in BEVDepth.
BEVFusion is single image based, no temporal module.
- Advantages of different sensors: Camera capture rich semantic information, lidar provide accurate spatial information, and radar offers instant velocity estimation.
- Projecting lidar to camera (RGBD data format) is a geometrically lossy process, as two neighbors on the depth map can be faraway from each other in 3D space. Projecting camera to lidar is a semantically lossy process, as point level fusion throws away the semantic density of camera features.
- 3D Object detection is a geometric-oriented task, map segmentation is semantic-oriented.
- Network acceleration. BEV voxel pooling is surprisingly slow.
- Grid association: The mapping between each camera pixel and BEV voxel is fixed (as the distance bin is also fixed), and therefore can be precomputed. [2X Speedup]
- Internal reduction: all points within the same BEV are aggregated with some symmetric function (mean, max, sum). LSS used the cumsum trick based integral image, but this way is not efficient. To accelerate this, BEVFusion directly assigns a GPU thread to each grid and process that cell. [20X Speedup]
- Voxel Pooling is also improved in BEVDepth and M2BEV.
- Fully convolutional fusion
- Camera BEV features may be misaligned with lidar BEV features. The convolutional layers in the BEV encoder can compensate for such local misalignments.
- Jointly training different tasks together (detection of dynamic objects and semantic segmentation of static objects) has a negative impact on the performance of each individual task ("negative transfer"). --> This is also observed in M2BEV and PETRv2.
- The night time performance of camera only performance (BEVFusion version) is worse than lidar-only performance (center point). --> This is a major risk of BEVNet in production.
- Questions and notes on how to improve/revise the current work