StreamPETR1: StreamPETR with EVA-02 Backbone

This repository contains the source/configuration needed to define the StreamPETR-style camera-only 3D detector with an EVA-02-Large backbone.

No model weights are included.

Included

  • StreamPETR EVA-02 config: projects/configs/StreamPETR/stream_petr_eva02_flash_800_bs2_seq_24e.py
  • RepDETR3D EVA-02 config: projects/configs/RepDETR3D/repdetr3d_eva02_800_bs2_seq_24e.py
  • Custom model/plugin code under projects/mmdet3d_plugin/
  • Local MMDetection3D Python/config support files under mmdetection3d/
  • Utility scripts needed by the configs and data conversion workflow

Not Included

  • Checkpoints or pretrained weights (.pth, .pt, .bin, .safetensors, etc.)
  • nuScenes data
  • Work directories, evaluation outputs, logs, or cache files

Model Details

  • Architecture: StreamPETR-style temporal 3D detector
  • Backbone: EVA-02 Large ViT, patch size 16, embed dim 1024, depth 24, 16 heads
  • Input: Multi-view camera images
  • Dataset: nuScenes
  • Task: Camera-only 3D object detection
  • Framework: MMDetection3D with a custom StreamPETR plugin

Users must provide their own EVA-02 and detector checkpoints locally.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support