Compatibility

Image semantic segmentation

• Datasets:
  • Rellis3D
  • GOOSE
  • Custom GAIA format: Parquet file containing samples and labels relative paths and a JSON file with the dataset ontology.
  • Generic: simply assumes a different directory per split, different suffixes for samples and labels, and a JSON file containing the dataset ontology.
• Models:
  • PyTorch (TorchScript compiled format and native modules):
    • Input shape: (batch, channels, height, width)
    • Output shape: (batch, classes, height, width)
  • Tensorflow (SavedModel compiled format and native Tensorflow/Keras modules):
    • Input shape: (batch, height, width, channels)
    • Output shape: (batch, height, width, classes)
    • JSON configuration file format:
  • ONNX: coming soon

  Each model must be coupled with a JSON configuration file:

    {
        "normalization": {
            "mean": [<r>, <g>, <b>],
            "std": [<r>, <g>, <b>]
        },
        "resize": {  # optional
            "width": <px>,
            "height": <px>
        },
        "crop": {  # optional
            "width": <px>,
            "height": <px>
        },
        "batch_size": <n>
    }
  

• Metrics:
  • Intersection over Union (IoU), Accuracy
• Computational cost:
  • Number of parameters, average inference time, model size

LiDAR semantic segmentation

• Datasets:
  • Rellis3D
  • GOOSE
  • Custom GAIA format: Parquet file containing samples and labels relative paths and a JSON file with the dataset ontology.
  • Generic: simply assumes a different directory per split, different suffixes for samples and labels, and a JSON file containing the dataset ontology.
• Models:
  • PyTorch (TorchScript compiled format and native modules). As of now, we have tested RandLA-Net and KPConv from Open3D-ML.
    • Input shape: defined by the input_format tag.
    • Output shape: (num_points)
    • JSON configuration file format:

      {
          "seed": 42,
          "input_format": "o3d_randlanet",
          "sampler": "spatially_regular",
          "recenter": {
              "dims": [
                  0,
                  1
              ]
          },
          "ignored_classes": [
              "void"
          ],
          "num_points": 45056,
          "grid_size": 0.06,
          "num_neighbors": 16,
          "sub_sampling_ratio": [
              4,
              4,
              4,
              4
          ]
      }
    

  • ONNX: coming soon
• Metrics:
  • Intersection over Union (IoU), Accuracy
• Computational cost:
  • Number of parameters, average inference time, model size

Object detection

Coming soon.