Add files using upload-large-folder tool

README.md

@@ -129,13 +129,32 @@ in this cleaned release. Use `manifest/frames_v1.parquet` and the identity
 manifests under `manifest/` as the source of truth for the v1 MOS/SOS release;
 use `manifest/frames_v2.parquet` when optional ego rows are present.
 
+This release applies a storage-only, lossless re-encoding to the published
+MOS/SOS data. Every transformed file was decoded immediately after writing and
+compared with its source array/value:
+
+* `rgb/` and `fisheye/` PNGs are stored as lossless WebP.
+* `depth/` and `sam2/masks/` remain PNG, re-encoded at compression level 9;
+  their 16-bit modes and pixel values are unchanged.
+* `cam_pose/` changes from per-frame `.npz` wrappers to per-frame `.npy`
+  `(7,) float64` vectors ordered as `[x, y, z, qx, qy, qz, qw]`.
+* `masks_aux` and `sam2_meta` are unchanged and retain the exact tar hashes
+  from the previous release.
+
+Across 1,451,790 packed members, logical storage falls from **248.5 GB** to
+**213.0 GB** (35.5 GB / 14.3% smaller) without changing decoded data. Optional
+ego metadata remains available, but no unpublished ego tar is added by this
+update.
+
 | modality | files | bytes |
 |---|---:|---:|
-| `cam_pose` | 110,000 | 163.3 MB |
-| `depth` | 110,000 | 16.9 GB |
-| `fisheye` | 220,000 | 63.4 GB |
-| `rgb` | 110,000 | 41.0 GB |
-| `masks` | 110,000 | 341.7 MB |
+| `cam_pose` | 110,000 | 20.2 MB |
+| `depth` | 110,000 | 13.2 GB |
+| `fisheye` | 220,000 | 56.9 GB |
+| `rgb` | 110,000 | 24.3 GB |
+| `masks` | 110,000 | 212.3 MB |
+| `masks_aux` | 790,631 | 118.4 GB |
+| `sam2_meta` | 1,159 | 535.6 kB |
 
 ## Quick start
 

manifest/current.json

@@ -78,5 +78,14 @@
     "mos_mask_object_map": "v1",
     "ego_frames": "v1",
     "mos_ego_mask_object_map": "v1"
+  },
+  "modality_extensions": {
+    "rgb": ".webp",
+    "depth": ".png",
+    "fisheye": ".webp",
+    "fisheye_left": ".webp",
+    "fisheye_right": ".webp",
+    "masks": ".png",
+    "cam_pose": ".npy"
   }
 }

rpx_croissant.json

@@ -44,12 +44,12 @@
   },
   "@type": "sc:Dataset",
   "name": "RPX",
-  "description": "RPX (Robot Perception X) is a unified real-world RGB-D benchmark for evaluating robot perception under embodied deployment conditions. It comprises 75,000 frames across 100 indoor scenes with ~70 daily object categories, captured under a three-phase reconfiguration protocol (clutter, human interaction, clean) using Intel D435 and T265 sensors. Ground-truth annotations include instance masks, tracklets, metric depth, 6-DoF camera pose, and language attributes, enabling evaluation of 8 perception tasks on identical scenes and objects.\n\nNote on on-disk shape: per-frame records are physically packed into per-(scene, phase, modality) tar shards on the HF repo (scenes/<scene>/<phase>/<modality>.tar for raw modalities; labels/<modality>/v<N>.tar for versioned labels). The rpx_benchmark.dataset_hub downloader resolves selective task+split downloads from these shards transparently.",
+  "description": "RPX (Robot Perception X) is a unified real-world RGB-D benchmark for evaluating robot perception under embodied deployment conditions. It comprises 110,000 released MOS/SOS frames across 100 indoor multi-object scenes and 70 selected single-object captures, using Intel D435 and T265 sensors. Ground-truth annotations include instance masks, tracklets, metric depth, 6-DoF camera pose, and language attributes, enabling evaluation of perception tasks on identical scenes and objects.\n\nNote on on-disk shape: per-frame records are physically packed into per-(scene, phase, modality) tar shards on the HF repo (scenes/<scene>/<phase>/<modality>.tar for raw modalities; labels/<modality>/v<N>.tar for versioned labels). RGB and fisheye frames use lossless WebP; depth and masks use lossless PNG; camera poses use per-frame NPY float64 vectors. Every transformed file was round-trip verified against its source. The rpx_benchmark.dataset_hub downloader resolves selective task+split downloads from these shards transparently.",
   "conformsTo": "http://mlcommons.org/croissant/1.0",
   "citeAs": "TODO: add bibtex citation after paper is published",
   "license": "https://creativecommons.org/licenses/by/4.0/",
   "url": "https://huggingface.co/datasets/IRVLUTD/RPX",
-  "version": "1.0.0",
+  "version": "1.1.0",
   "isLiveDataset": false,
   "creator": {
     "@type": "Organization",
@@ -73,7 +73,7 @@
       "@type": "cr:FileObject",
       "@id": "rpx-full",
       "name": "rpx-full",
-      "description": "Full 75,000-frame dataset with all annotations",
+      "description": "Full 110,000-frame MOS/SOS dataset with all annotations",
       "contentUrl": "https://anonymous.4open.science/r/RPX/full",
       "encodingFormat": "application/zip",
       "sha256": "TODO"
@@ -126,14 +126,14 @@
           "@type": "cr:Field",
           "@id": "frames/rgb_path",
           "name": "rgb_path",
-          "description": "Relative path to RGB image (PNG, 1920x1080)",
+          "description": "Relative path to RGB image (lossless WebP, RGB uint8)",
           "dataType": "sc:Text"
         },
         {
           "@type": "cr:Field",
           "@id": "frames/depth_path",
           "name": "depth_path",
-          "description": "Relative path to depth map (NPY float32, metres)",
+          "description": "Relative path to depth map (lossless PNG, uint16 millimetres)",
           "dataType": "sc:Text"
         },
         {
@@ -147,7 +147,7 @@
           "@type": "cr:Field",
           "@id": "frames/camera_pose",
           "name": "camera_pose",
-          "description": "4x4 SE(3) camera-to-world matrix from T265 (float64)",
+          "description": "Relative path to a (7,) float64 NPY pose vector [x, y, z, qx, qy, qz, qw] from T265",
           "dataType": "sc:Text"
         },
         {
@@ -271,4 +271,4 @@
       ]
     }
   ]
-}
+}