本文主要是记录地平线官方提供的可在J5芯片上部署的参考算法:
参考算法 | 数据集 | FPS |
---|---|---|
PointPillars | KITTI | 116 (双核) |
CenterPoint | Nuscenes | 98.72(双核) |
Lidar MultiTask | Nuscenes | 98.72(双核) |
FCOS3D | Nuscenes | 589 (双核) |
PETR | Nuscenes | 8.3(双核) |
bev_mt_lss | Nuscenes | 138(单核) |
bev_mt_ipm | Nuscenes | 106(单核) |
bev_mt_ipm_4d | Nuscenes | 213(双核) |
bev_mt_gkt | Nuscenes | 42(单核) |
bev_cft | Nuscenes | 35.11(双核) |
双目深度估计 | SceneFlow | 249.68(双核) |
DETR | COCO | 62.29(双核) |
GANet(车道线检测) | CULane | 2431(双核) |
MOTR(多目标跟踪) | mot17 | 73.98(双核) |
关键点检测 | CarFusion | 3290.12(双核) |
DenseTNT | Argoverse 1 | 1826(双核) |
MixVarGENet | ImageNet | 5837(双核) |
Swin Transformer | ImageNet | 133(双核) |
地平线官方提供的可在J6芯片上部署的参考算法:
参考算法 | 数据集 | J6E FPS | J6M FPS |
---|---|---|---|
FlashOcc | Nuscenes | 68.6 | 96.23 |
QCNet | Argoverse2 | 35.80 | 208.39 |
Sparse4D | Nuscenes | 29.44 / 62 | - |
BEVFormer | Nuscenes | 20.48 | - |
MapTR | Nuscenes | 26.68 / 67.54 | - |
BEVFusion | Nuscenes | 30.96 | - |
J5参考算法
1 PointPillars(Kitti)
算法介绍: https://developer.horizon.ai/forumDetail/118364000835765874
2 Centerpoint (Nuscenes)
算法介绍:https://developer.horizon.ai/forumDetail/163807121354434630
3 Lidar MultiTask (Nuscenes)
算法介绍: https://developer.horizon.cc/forumDetail/173442588425246842
4 FCOS3D (Nuscenes)
算法介绍:https://developer.horizon.ai/forumDetail/146177165367615529
5 PETR (Nuscenes)
算法介绍: https://developer.horizon.ai/forumDetail/146177165367615527
6 BEV参考算法 (Nuscenes)
7 双目深度估计(SceneFlow)
算法介绍:https://developer.horizon.cc/forumDetail/189779523032809474
8 DETR(COCO)
算法介绍:https://developer.horizon.cc/forumDetail/191823650440318976
9 GANet (CULane)
算法介绍:https://developer.horizon.ai/forumDetail/163807121354434982
10 MOTR (mot17)
算法介绍:https://developer.horizon.cc/forumDetail/174216017545977856
11 关键点检测(CarFusion)
算法介绍: https://developer.horizon.cc/forumDetail/174216054053200930
12 DenseTNT (Argoverse 1)
算法介绍: https://developer.horizon.cc/forumDetail/174216054053200932
13 MixVarGENet (ImageNet)
算法介绍: https://developer.horizon.ai/forumDetail/143772473308124160
14 SwinTransformer (ImageNet)
算法介绍:https://developer.horizon.ai/forumDetail/143772473308124166
J6参考算法
1. FlashOcc(Nuscenes)
2. QCNet(Argoverse2)
算法介绍(V2.0):https://developer.horizon.auto/blog/10004
3. Sparse4D(Nuscenes)
算法介绍(V2.0):https://developer.horizon.auto/blog/12622
模型 | 数据集 | Input shape | backbone | num_anchor | NDS | mAP | FPS | Latency |
---|---|---|---|---|---|---|---|---|
SparseBevOE | Nuscenes | 6x3x256x704 | HENET | 384 | 0.5424/0.53 | 0.4040 | 62 | 13.28ms |
4. BEVFormer(Nuscenes)
算法介绍 (V1.0):https://auto-developer.horizon.cc/developerForum?fullPath=/home/community/bbsdetail?bid=617255194813960192
算法介绍 (V2.0):https://developer.horizon.auto/blog/10006
5. MapTR(Nuscenes)
算法介绍(V2.0):https://developer.horizon.auto/blog/10142
6. BEVFusion(Nuscenes)
算法介绍:https://developer.horizon.auto/blog/10003