Dataset Viewer
Auto-converted to Parquet Duplicate
chunks
list
config
dict
updated_at
string
[ { "chunk_bytes": 236041154, "chunk_size": 50, "dim": null, "filename": "chunk-0-0.bin" }, { "chunk_bytes": 240768367, "chunk_size": 51, "dim": null, "filename": "chunk-1-0.bin" }, { "chunk_bytes": 240766803, "chunk_size": 51, "dim": null, "filename": "chunk-10...
{ "chunk_bytes": 256000000, "chunk_size": null, "compression": null, "data_format": [ "numpy", "numpy", "no_header_numpy:8", "numpy", "pickle", "numpy", "numpy", "no_header_numpy:9", "pickle", "pickle" ], "data_spec": "[1, {\"type\": \"builtins.dict\", \"context\": \"...
1776241568.6028094

Mars HiRISE DTM — LitData Streaming Dataset

Pre-processed Mars HiRISE orthoimage + DTM patches in LitData optimized streaming format.

Quick Start

from litdata import StreamingDataset

# Stream directly from HuggingFace — no full download needed
train_ds = StreamingDataset(input_dir="hf://datasets/SuperComputer/mars_hirise_dtm_processed-d737fb4b0c75695d/train")
val_ds   = StreamingDataset(input_dir="hf://datasets/SuperComputer/mars_hirise_dtm_processed-d737fb4b0c75695d/val")
test_ds  = StreamingDataset(input_dir="hf://datasets/SuperComputer/mars_hirise_dtm_processed-d737fb4b0c75695d/test")

sample = train_ds[0]

# Core Tensors
print(sample["image"].shape)        # (3, H, W) float16
print(sample["dtm"].shape)          # (3, H, W) float16
print(sample["confidence"].shape)   # (1, H, W) float16

# Physical / Lighting Parameters
print(sample["sun_vector"].shape)   # (3,)      float32
print(sample["intensity"])          # float32
print(sample["ambient"])            # float32

# Original / Unnormalized Data
print(sample["original_image"].shape) # (C, H, W) float16
print(sample["original_dtm"].shape)   # (1, H, W) float16
print(sample["trend_params"].shape)   # (3,)      float16

# Normalization / Processing Metadata
print(sample["residual_scale"])     # float32
print(sample["raw_residual_p98"])   # float32
print(sample["key"])                # str (e.g., "left_red")

Fields

During extraction, model inputs are quantized to float16 to optimize streaming bandwidth. Physical lighting parameters remain float32.

Key Dtype Shape Description
image float16 (3, H, W) Normalized HiRISE orthoimage [-1, 1]
dtm float16 (3, H, W) Normalized DTM elevation
confidence float16 (1, H, W) Binary valid data mask (eroded/cleaned)
sun_vector float32 (3,) Estimated sun direction (OLS, unit-normalized)
intensity float32 scalar Estimated sun intensity
ambient float32 scalar Estimated ambient light
original_image float16 (C, H, W) Unnormalized resized orthoimage
original_dtm float16 (1, H, W) Unnormalized resized DTM elevation
trend_params float16 (3,) LSQR detrend parameters for the DTM plane
residual_scale float32 scalar Normalization scale applied to the DTM residual
raw_residual_p98 float32 scalar 98th percentile of the raw topographic residual
key string scalar Orthoimage source used (e.g., left_red)

Preprocessing Configuration

Config Hash: d737fb4b0c75695d

This dataset was generated with the following pipeline parameters:

hirise:
  root: /scratch/mars_hirise_dtm
  include_ortho: true
  ortho_type: RED
  ortho_scale: null
  download: false
  reuse_cache: true
  target: null
  bbox:
  - 60
  - 0
  - 90
  - 30
sampler:
  size: 0.018
  length: null
resolution: 512
dtm_normalization: relative
Downloads last month
125