Initial snapshot: hand motion pipeline (Dyn-HaMR + dex-retargeting + SPIDER)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

third_party/vipe/configs/default.yaml

@@ -0,0 +1,16 @@
+defaults:
+  - _self_
+  - streams: raw_mp4_stream
+  - pipeline: ???
+  - override hydra/hydra_logging: none
+  - override hydra/job_logging: disabled
+
+ray: false
+prefilter: false
+
+hydra:
+  output_subdir: null
+  job:
+    chdir: false
+  run:
+    dir: .

third_party/vipe/configs/pipeline/dav3.yaml

@@ -0,0 +1,8 @@
+defaults:
+  - default
+
+slam:
+  keyframe_depth: dav3
+
+post:
+  depth_align_model: "mvd_dav3"

third_party/vipe/configs/pipeline/default.yaml

@@ -0,0 +1,47 @@
+defaults:
+  - /slam: default
+
+instance: vipe.pipeline.default.DefaultAnnotationPipeline
+
+# Init configs
+init:
+  camera_type: "pinhole"
+  intrinsics: "geocalib"
+  instance:
+    kf_gap_sec: 2.0
+    phrases: 
+      - person
+      - animal
+      - vehicle
+      - ball
+      - balloon
+      - gun
+      - pet
+      - car
+      - bus
+    add_sky: true
+
+slam:
+  keyframe_depth: unidepth-l
+  optimize_intrinsics: ${neq:${..init.intrinsics},"gt"}
+
+# Post-processing configs
+post:
+  depth_align_model: "adaptive_unidepth-l_svda"
+
+# Output configs
+output:
+  # Path to save results
+  path: vipe_results/
+  skip_exists: false
+
+  # Save artifacts and a info file
+  save_artifacts: false
+
+  # Save reconstruction from SLAM
+  save_slam_map: false
+
+  # Visualization videos to BASE_PATH/vipe/xxx.mp4
+  save_viz: true
+  viz_downsample: 2
+  viz_attributes: [['rgb', 'instance'], ['depth', 'pcd']]

third_party/vipe/configs/pipeline/dynhamr_cameras.yaml

@@ -0,0 +1,9 @@
+# Dyn-HaMR consumes camera poses and intrinsics, not dense depth artifacts.
+defaults:
+  - default
+
+post:
+  depth_align_model: null
+
+output:
+  cameras_only: true

third_party/vipe/configs/pipeline/lyra.yaml

@@ -0,0 +1,8 @@
+defaults:
+  - default
+
+slam:
+  keyframe_depth: moge
+
+post:
+  depth_align_model: "adaptive_moge_vda"

third_party/vipe/configs/pipeline/no_vda.yaml

@@ -0,0 +1,5 @@
+defaults:
+  - default
+
+post:
+  depth_align_model: "adaptive_unidepth-l"

third_party/vipe/configs/pipeline/static_vda.yaml

@@ -0,0 +1,8 @@
+defaults:
+  - default
+
+init:
+  instance: null
+
+post:
+  depth_align_model: "adaptive_unidepth-l_vda"

third_party/vipe/configs/pipeline/wide_angle.yaml

@@ -0,0 +1,14 @@
+defaults:
+  - default
+
+init:
+  camera_type: "mei"
+
+slam:
+  keyframe_depth: unidepth-l
+
+post:
+  depth_align_model: null
+
+output:
+  viz_attributes: [['rgb', 'instance'], ['rectified', 'pcd']]

third_party/vipe/configs/slam/default.yaml

@@ -0,0 +1,53 @@
+buffer: 1024
+# weight for translation / rotation components of flow
+beta: 0.3
+# how much motion before considering new keyframe
+filter_thresh: 2.4
+# number of warmup frames
+warmup: 8
+# threshold to create a new keyframe
+keyframe_thresh: 4.0
+# add edges between frames whithin this distance
+frontend_thresh: 16.0
+# frontend optimization window
+frontend_window: 25
+# force edges between frames within radius
+frontend_radius: 2
+# non-maximal supression of edges
+frontend_nms: 1
+# Whether to initialize sequentially
+seq_init: true
+# Keyframe ids for running backends
+frontend_backend_iters: [16, 64, 256]
+backend_thresh: 22.0
+backend_radius: 2
+backend_nms: 3
+backend_iters: 24
+# Initial values for disparity
+init_disp: 1.0
+# Whether to optimize intrinsics
+optimize_intrinsics: ???
+# Whether to optimize rig
+optimize_rig_rotation: false
+# Whether add reproj factors for cross-views
+cross_view: true
+cross_view_idx: null  # null means default
+adaptive_cross_view: false  # Backend only, recompute cross_view_idx
+# Chunk size of the infill
+infill_chunk_size: 16
+# Whether or not to infill dense disparity
+infill_dense_disp: False
+# Filter threshold for SLAM map (percentage of the mean depth difference allowed)
+# (This is also used for dense disp extraction)
+map_filter_thresh: 0.05
+# Whether to visualize the results using rerun
+visualize: false
+# Use keyframe depth for recovering metric depth
+keyframe_depth: metric3d-small
+
+ba:
+  dense_disp_alpha: 0.001
+  intrinsics_damping_scale: 1.0
+
+sparse_tracks:
+  name: dummy

third_party/vipe/configs/streams/frame_dir_stream.yaml

@@ -0,0 +1,7 @@
+instance: vipe.streams.frame_dir_stream.FrameDirStreamList
+
+base_path: ??
+frame_start: 0
+frame_end: -1
+frame_skip: 1
+cached: false

third_party/vipe/configs/streams/raw_mp4_stream.yaml

@@ -0,0 +1,7 @@
+instance: vipe.streams.raw_mp4_stream.RawMP4StreamList
+
+base_path: ??
+frame_start: 0
+frame_end: 1000
+frame_skip: 1
+cached: false