qizhangslam/e2e-stream-slam-training-dataset

e2e-stream-slam training assets

Reproducibility bundle for the V4 SLAMFormer ablation suite.

Code: https://github.com/SlamMate/e2e-semantic-SLAM/tree/submap (commit 3195a7a)

Contents

Checkpoints

File	Size	Role
checkpoints/v1_paper_ckpt10.pth	3.6 GB	SLAMFormer paper base ckpt (10 ep on the paper datasets). PRETRAINED init for V3 Scale Token training.
checkpoints/v3_scale_token_ckpt2.pth	3.8 GB	V3 Scale Token epoch-2 (3 ep, 3×A6000, FSDP). PRETRAINED init for V4-B'' and V4-final.
checkpoints/v4bpp_ckpt1.pth	8.6 GB	Headline result: V4-B'' epoch-1 = V3 ckpt-2 → FT{paper + scannetpp + mvs_synth + vKITTI2}. Eval: 41.3 m KITTI / 0.160 TUM (V4-oracle chain-TRS(6,4)).

Configs (Hydra snapshots)

File	Role
configs/v3_scale_token_hydra.yaml	Training config captured for v3_scale_token_ckpt2
configs/v4bpp_hydra.yaml	Training config captured for v4bpp_ckpt1

Training chain

v1_paper_ckpt10.pth
  └── [sbatch_scale_token_v3_fsdp_3gpu.sh] ──▶ v3_scale_token_ckpt2.pth
                                                  └── [sbatch_v4bpp_v3base_vkitti_fsdp_3gpu.sh] ──▶ v4bpp_ckpt1.pth

All sbatches live in the code repo under slam/sbatch_*.sh.

Currently running (not yet in this bundle, will be added when done):

• sbatch_v4_final_3gpu.sh — V4 final recipe, 3 epochs
• sbatch_v4_novk_3gpu.sh — V4 noVK ablation, 3 epochs

How to download

pip install huggingface_hub
huggingface-cli download qizhangslam/e2e-stream-slam-training-dataset \
    checkpoints/v4bpp_ckpt1.pth \
    --repo-type dataset \
    --local-dir ./ckpt

How to evaluate

After cloning the code repo and placing eval datasets under data/eval/{kitti_odometry,tum}/, the provided run_v4bpp_e1_*.sh scripts reproduce the §15.12 / §15.13 / §15.14 / §15.16 ablation numbers from the downloaded v4bpp_ckpt1.pth.

Dataset requirements (NOT in this HF repo)

Reproducing training from scratch additionally requires the processed training datasets:

• data/train/processed_scannetpp/
• data/train/processed_mvs_synth/
• data/train/processed_vkitti/

And for evaluation:

• data/eval/kitti_odometry/dataset/sequences/{00..10}/
• data/eval/tum/rgbd_dataset_freiburg1_*/

See the upstream SLAMFormer paper and our code repo's CHANGELOG for dataset preparation details.