Datasets:
TactileNet-A-classification
Private evaluation data for frozen tactile encoders: 480 complete episodes, 620,678 frames, 8 materials and 32 specifications. Each specification has 15 episodes. The fixed train/validation/test split contains 288/96/96 episodes (9/3/3 per specification).
评测代码和接入教程:B0B8K1ng/TactileNet。用自己的编码器及原生归一化提取完整片段特征,按时间顺序拼接,再用统一逻辑回归评测。代码仓库同样为私有,需获授权后访问。
Files
TactileNet-A-classification.tar.gz: 223,194,837 bytes (212.86 MiB), the recommended evaluation download. All original raw tactile values from both hands and aligned timestamps are preserved losslessly, along with sensor attributes and task SOP metadata. No cropping, frame subsampling, normalization or phase selection. Camera/image/joint payloads are omitted from this small package.manifest.json,split.json,classes.json: portable episode inventory, frozen split and material/specification hierarchy. Also included inside the archive.CHECKSUMS.sha256,ARCHIVE_SHA256SUMS: HDF5 and archive checksums.data_audit.json,archive_audit.json: complete source-array/metadata and archive verification records.
This release contains the complete tactile time-series evaluation package only. It extracts to simple material/specification directories: TactileNet-A-classification/data/材料/规格/原文件名.hdf5. Camera/video, joint data and separate PNG inspection previews are not distributed.
Archive SHA256: ab80a9ec1f66b90b7229dc8362e856c73079d3fc6d01a4c65b41ce60aa229644.
Sensor schema
Within each HDF5:
| Key | Shape / type |
|---|---|
dataset/observation/aligned_timestamp |
[T], int64, milliseconds |
dataset/observation/lefthand/tactile/data |
[T,3465], float32 |
dataset/observation/righthand/tactile/data |
[T,3465], float32 |
Each tactile dataset has sensor_names and sensor_lengths attributes. Each sensor segment starts with 30 header columns, followed by XYZ values. Removing these headers yields 1005 XYZ taxels per hand. Do not directly reshape all 3465 columns into XYZ. The evaluation code supplies a validated loader returning [T,2,1005,3] without normalization. Zero-valued frames are retained.
Scope and protocol
- Excludes the legacy 60-episode plastic set and five
diff_gloveepisodes. All 480 current A-group episodes are included exactly once. - All time phases are retained; this is not the index-finger sliding-only subset.
- The original fixed split is preserved. Recording groups appear in all subsets; results measure recording-level generalization within observed groups, not held-out participants or gloves.
- Each small class has only three test episodes. Interpret per-class F1 at that sample size.
- Actual ceramic labels include
条纹浮雕砖;无釉粗陶片does not occur and has not been substituted. - Labels, episode IDs, absolute timestamps and filenames must not be supplied as classifier features. Use the tactile signal and the encoder's own frozen preprocessing.
- Input normalization belongs to the encoder. The classifier's variance filter and feature scaler are fitted on training data only. Validation selects regularization; test is used after selection.
- Full-sequence concatenation with explicit tail handling is the current protocol. It differs from older internal mean-pooled window experiments; their scores are not interchangeable.
This is a private release. No public redistribution license is assigned in this dataset card.
Mirrors: Hugging Face · ModelScope.
- Downloads last month
- 20