smerkd/arrow-presentation-videos

ARROW presentation videos (TUM RGB-D fr3/walking_xyz)

Renders of an ARROW checkpoint (Stage2 v2, step 3000 of 6000) on TUM RGB-D rgbd_dataset_freiburg3_walking_xyz, which is held out from training. Panels are stacked side by side with no text.

File	Panels, left to right
arrow_reconstruction_features_tracks.mp4	input RGB · ARROW pointmap · deferred-renderer RGB · EUPE feature PCA · EUPE cosine-similarity heatmap to the people · ARROW 2D tracks
arrow_edit_remove_floor_chair_preview.mp4	orbit: deferred RGB · floor removed · chair removed (EUPE cosine; short preview, full version to follow)
arrow_slam_delta_decoding_third_person.mp4	third-person view of ARROW-SLAM with delta decoding: the map grows; newly decoded points flash green and fade, reused points keep their colour

Gaussians: one per pixel (stride 1), built by the training code's own Gaussian head and rendered with the deferred renderer. People / floor / chair seeds are chosen with the SigLIP2 teacher on one frame; every heatmap and removal uses the model's own splatted EUPE features (centred cosine). The delta-decoding video uses the released ARROW backbone.

refiner_s4250/ (Stage2 v2 + joint map refiner, step 4250; every pixel, refiner applied)

File	Panels, left to right
reconstruction_features_tracks.mp4	input · ARROW pointmap · deferred RGB · EUPE PCA · EUPE cosine heatmap (people) · long-term ARROW tracks (re-queried into the next chunk, persistent colours)
siglip_query_removal_camera_path.mp4	input · deferred RGB · floor removed · person removed. Removal = the model's own SigLIP text query (closed vocabulary) per Gaussian; holes are filled from frames +-8 where the background was visible
siglip_query_removal_topdown_orbit.mp4	top-down sweep: deferred RGB · floor removed · person removed
stills/frameNNN_all_queries.png	input · deferred RGB · SigLIP concept probability for person, floor, computer monitor, chair, desk, keyboard, trash can, plant
stills/per_query/	input next to each single query heatmap

Every pixel becomes a Gaussian (ARROW geometry where the depth sensor has none). Known limits: SigLIP queries are reliable for person / floor / computer monitor and weak for rarer concepts; parts of a moving person not labelled "person" in some frames leave ghosts after removal.

arrow_slam_delta_decoding_full_map_topdown.mp4

Third-person, top-down view of ARROW-SLAM with delta decoding, all decoded points (not only the confident static map).

slam_s4250/ — ARROW-SLAM on the stage2_v2_refiner step-4250 checkpoint

The checkpoint's ARROW backbone (exported to the release format) runs the default delta-decoding ARROW-SLAM; the Gaussians come from the same checkpoint's heads, and every frame is placed and viewed with ARROW-SLAM's estimated pose (GT only fixes the global frame and metric scale, Sim(3) on camera centres). 225 frames (15 s at 15 fps), one Gaussian per pixel, joint map refiner applied. One folder per scene: walking_xyz, walking_halfsphere (TUM fr3), bonn_crowd, bonn_person_tracking (Bonn RGB-D Dynamic).

file	panels (left to right)
slam_map_trajectory.mp4	input · whole Gaussian map from the SLAM camera · third-person follow view with trajectory + frustum · top-down overview. The map grows over time; people appear only at the current time; outside views cut the ceiling
reconstruction_features_tracks.mp4	input · pointmap · deferred RGB · EUPE PCA · EUPE cosine heatmap (people) · long-term tracks
siglip_query_removal_camera_path.mp4	input · deferred RGB · floor removed · person removed (SigLIP text query)
siglip_query_removal_orbit.mp4	top-down sweep: deferred RGB · floor removed · person removed
delta_gaussians_tracks_third_person.mp4	Gaussian delta decoding, third-person follow view: input with the person query points · deferred RGB · EUPE PCA of the splatted features. The map starts from the first frame and each frame adds only new Gaussians (pixels the current map does not explain), shown green and fading back over 6 frames; people are rendered at the current time only; dense ARROW 3D track trails of every person (6 px query spacing); smoothed SLAM trajectory + current frustum
overhead_person.mp4 (bonn_person_tracking)	third-person overhead view following the camera: input with the person query points · person removed · splatted EUPE features (PCA) · RGB + person cosine heatmap; ARROW 3D track trails of the person query points (re-queried into the next chunk, metric via sensor depth, placed with the SLAM poses) on the rendered panels
delta_decoding_third_person_topdown.mp4	third-person delta decoding over the full sequence: newly decoded points flash green and fade
stills/frameNNN_all_queries.png	input · deferred RGB · SigLIP concept probability per query

SLAM trajectory error (all frames, Sim(3) aligned, coverage = tracked frames / all frames):

scene	released ARROW	stage2_v2_refiner s4250	coverage (released / new)
walking_xyz	1.71 cm	2.36 cm	96 % / 97 %
walking_halfsphere	2.83 cm	2.86 cm	93 % / 94 %
bonn_crowd	1.52 cm	1.40 cm	98 % / 98 %
bonn_person_tracking	2.85 cm	4.03 cm	93 % / 93 %

Removal of static classes (floor) is a multi-view vote: a Gaussian is removed when any frame of the set sees it (depth-tested) as that class, not only its own frame; people are removed per current frame.

Some surfaces in the SLAM-camera view are doubled where neighbouring windows placed with SLAM poses disagree slightly (real map inconsistency, not hidden).

TUM RGB-D: Sturm et al., IROS 2012 (CC BY 4.0). Bonn RGB-D Dynamic: Palazzolo et al., IROS 2019.