Datasets

Please follow MonST3R and Spann3R to download the datasets used by the current evaluation scripts: ScanNet, TUM-dynamics, Bonn, 7Scenes, and NRGBD.

ScanNet

To prepare the ScanNet dataset, execute:

python datasets_preprocess/long_prepare_scannet.py # You may need to change the path of the dataset

TUM-dynamics

To prepare the TUM-dynamics dataset, execute:

python datasets_preprocess/long_prepare_tum.py # You may need to change the path of the dataset

The relpose and video_depth evaluators can also read the original TUM-RGBD layout directly: ROOT/rgbd_dataset_*/rgb, ROOT/rgbd_dataset_*/depth, rgb.txt, depth.txt, and groundtruth.txt.

Bonn

To prepare the Bonn dataset, execute:

python datasets_preprocess/long_prepare_bonn.py # You may need to change the path of the dataset

The relpose and mv_recon evaluators can also read the original Bonn layout directly: ROOT/rgbd_bonn_*/rgb, ROOT/rgbd_bonn_*/depth, rgb.txt, depth.txt, and groundtruth.txt.

7Scenes and NRGBD

The 3D reconstruction evaluator follows the 7Scenes/NRGBD layout used by S-VGGT:

• 7Scenes: ROOT/<scene>/TrainSplit.txt, ROOT/<scene>/TestSplit.txt, and ROOT/<scene>/seq-xx/frame-000000.(color.png|depth.proj.png|pose.txt).
• NRGBD: ROOT/<scene>/images/img0.png, ROOT/<scene>/depth/depth0.png, and ROOT/<scene>/poses.txt.

You can pass roots explicitly with --seven_scenes_root and --nrgbd_root, or set EVAL_7SCENES_ROOT and EVAL_NRGBD_ROOT.

Evaluation Scripts

Results will be saved in eval_results/*.

Camera Pose Estimation

CUDA_VISIBLE_DEVICES=6,7 bash eval/relpose/run_scannet.sh # You may need to change [--num_processes] to the number of your gpus
CUDA_VISIBLE_DEVICES=6,7 bash eval/relpose/run_tum.sh # Uses original TUM-RGBD layout by default; set EVAL_TUM_DATASET=tum_s1_1000 for prepared data

For TUM-dynamics pose evaluation on the original dataset layout, pass the dataset root that contains the rgbd_dataset_* sequence folders:

python eval/relpose/launch.py \
  --weights src/cut3r_512_dpt_4_64.pth \
  --output_dir eval_results/relpose/tum/recal3r \
  --eval_dataset tum \
  --dataset_path path/to/tum \
  --pose_eval_stride 1 \
  --max_frames 1000 \
  --size 512 \
  --model_update_type recal3r

Use --seq_list rgbd_dataset_freiburg3_walking_xyz to evaluate only selected TUM sequences. The helper script also accepts EVAL_TUM_ROOT, EVAL_TUM_SEQ_LIST, EVAL_TUM_MAX_FRAMES, and EVAL_TUM_STRIDE.

For Bonn pose evaluation on the original dataset layout, pass the dataset root that contains the rgbd_bonn_* sequence folders:

python eval/relpose/launch.py \
  --weights src/cut3r_512_dpt_4_64.pth \
  --output_dir eval_results/relpose/bonn/recal3r \
  --eval_dataset bonn \
  --dataset_path path/to/rgbd_bonn_dataset \
  --pose_eval_stride 1 \
  --max_frames 500 \
  --size 512 \
  --model_update_type recal3r

Use --seq_list balloon2 crowd2 to evaluate only selected Bonn sequences. Sequence names may omit the rgbd_bonn_ prefix.

3D Reconstruction

CUDA_VISIBLE_DEVICES=5 bash eval/mv_recon/run.sh # Evaluates 7Scenes and NRGBD; set EVAL_7SCENES_ROOT and EVAL_NRGBD_ROOT when needed

The 3D reconstruction evaluator follows the 7Scenes/NRGBD layout used by S-VGGT. For direct evaluation, pass the dataset root with --dataset_root, or use --seven_scenes_root and --nrgbd_root:

python eval/mv_recon/launch.py \
  --weights src/cut3r_512_dpt_4_64.pth \
  --model_name recal3r \
  --output_dir eval_results/video_recon/7scenes_200/recal3r \
  --eval_dataset 7scenes \
  --dataset_root path/to/7-scenes \
  --max_frames 200 \
  --size 512 \
  --model_update_type recal3r

Video Depth

CUDA_VISIBLE_DEVICES=5 bash eval/video_depth/run_bonn.sh # You may need to change [--num_processes] to the number of your gpus
CUDA_VISIBLE_DEVICES=5 bash eval/video_depth/run_tum.sh # Uses original TUM-RGBD layout by default; set EVAL_TUM_DATASET=tum_s1_1000 for prepared data

For TUM-dynamics depth evaluation on the original dataset layout:

python eval/video_depth/launch.py \
  --weights src/cut3r_512_dpt_4_64.pth \
  --output_dir eval_results/video_depth/tum/recal3r \
  --eval_dataset tum \
  --dataset_path path/to/tum \
  --pose_eval_stride 1 \
  --max_frames 1000 \
  --size 512 \
  --model_update_type recal3r

python eval/video_depth/eval_depth.py \
  --output_dir eval_results/video_depth/tum/recal3r \
  --eval_dataset tum \
  --dataset_path path/to/tum \
  --pose_eval_stride 1 \
  --max_frames 1000 \
  --tum_depth_scale 5000 \
  --align scale\&shift

Use --seq_list rgbd_dataset_freiburg3_walking_xyz on both commands to evaluate selected TUM sequences.

For Bonn depth evaluation on the original dataset layout:

python eval/video_depth/launch.py \
  --weights src/cut3r_512_dpt_4_64.pth \
  --output_dir eval_results/video_depth/bonn/recal3r \
  --eval_dataset bonn \
  --dataset_path path/to/rgbd_bonn_dataset \
  --pose_eval_stride 1 \
  --max_frames 1000 \
  --size 512 \
  --model_update_type recal3r

python eval/video_depth/eval_depth.py \
  --output_dir eval_results/video_depth/bonn/recal3r \
  --eval_dataset bonn \
  --dataset_path path/to/rgbd_bonn_dataset \
  --pose_eval_stride 1 \
  --max_frames 1000 \
  --align scale\&shift

Use --seq_list balloon crowd2 on both commands to evaluate selected Bonn sequences.