EvoMem3D: baseline and final DETR3D model
Weights for EvoMem3D: Adaptive-Depth Temporal Representation Memory for Multi-Camera 3D Detection.
This repository contains exactly two model checkpoints, both using the EVA-02-L image backbone and the RepDETR3D detector implementation.
| Model | Weights | Matching config | Recorded result |
|---|---|---|---|
| Full-depth detector baseline | Baseline checkpoint | Config | Detector initialization baseline |
| Final EvoMem3D temporal detector | Final checkpoint | Config | 61.50 NDS / 52.76 mAP |
The final run's validation log records NDS 0.61505 and mAP 0.52759
on nuScenes validation. Metrics and provenance
identify that recorded evaluation. This is the motion-reset, feature-noise
0.10 temporal model from iter_42192.pth, also used for the project-page
rollouts. The baseline is the EVA-02 bs4 trained detector used to initialize
temporal training.
Download
from huggingface_hub import hf_hub_download
repo = "Kimhi/evomed3d-repdetr3d"
baseline = hf_hub_download(repo, "baseline/repdetr3d_eva02_baseline.pth")
final_model = hf_hub_download(repo, "evomem3d/evomem3d_repdetr3d_nds61_50.pth")
baseline_config = hf_hub_download(repo, "baseline/config.py")
final_config = hf_hub_download(repo, "evomem3d/config.py")
Both files preserve the original training snapshots, including optimizer state. The final checkpoint includes the backbone, detector heads, partial-feature projectors, memory updater, and trained depth policy. Model construction and evaluation require compatible EvoMem3D/RepDETR3D code. Load the full checkpoint after constructing the matching model so the trained policy is restored.
The configs are the expanded configs stored inside the corresponding checkpoints. For inference, disable pretrained initialization before loading the full checkpoint. The final model uses previous detector predictions for its teacher-forcing feature path during evaluation.
Verification
SHA-256 checksums match both released files to their local source artifacts. Checkpoint tensors and trained behavior are unchanged. In this implementation, internal partial routes L8/L12/L16 use inclusive zero-based stop indices and execute 9/13/17 blocks; Full executes 24.