Skip to content

Question about inference with object not in pixieverse dataset #15

@Li-Hao-yuan

Description

@Li-Hao-yuan

Background

I am currently using Pixie to predict object parameters for objects from the Objaverse dataset. While the framework work for the provided pixieverse data, I encountered a bottleneck when attempting to run the pipeline on new, unseen objects (those not included in the pre-processed pixieverse set).

The Problem

When transitioning from voxelization to dataset construction for a new object, the system fails to build the MaterialVoxelDataset due to a missing file: obj_id/sample_0/material_grid.npy

According to my understanding, for inference on novel objects, Pixie should ideally only require multi-view images. However, the MaterialVoxelDataset logic seems to expect a pre-existing material grid.

Current Workaround

After experimenting with the pipeline, I found that the issue can be bypassed by treating these Objaverse objects similarly to "real-world scenes." Specifically, I modified the logic in the pipeline by adding format_real_scene_sample before the neural segmentation step:

# Original logic
# sample_output_dir = generate_neural_segmentation(cfg, render_output_dir, paths)

# Modified logic
format_real_scene_sample(cfg, paths) # Manually trigger the creation of an empty/placeholder material_grid
sample_output_dir = generate_neural_segmentation(cfg, render_output_dir, paths)

By adding this step, the pipeline generates the necessary (though potentially empty) material_grid.npy, allowing the data flow to complete and the inference to proceed.

Therefore, my question is that is this the intended or standard way to handle out-of-dataset objects from Objaverse?

Thank you for your time and for this outstanding framework!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions