Learning for 3D Vision: Assignment IV


Point Cloud Processing

CMU 16825 Learning for 3D Vision

Tushar Nayak [tusharn]

1. Classification

resp
Prediction: Correct
resp
Prediction: Correct
resp
Prediction: Correct
resp
Prediction: Incorrect (Predicted as Chair, probably due to the shape of the base resembling the bases of some rotating chairs in the dataset)| Comments:
resp
Prediction: Incorrect (Predicted as Vase, probably due to the vague conical shape resembling an inverted vase)| Comments:
resp
Prediction: Incorrect (Predicted as Vase, probably due to the vague conical shape resembling an inverted vase)| Comments:

Accuracy on Test Set: 0.9717

Interpretation & Explination of Results:

2. Segmentation

Figure 1 (0.9915)

resp
Ground Truth
resp
Prediction

Figure 2 (0.9974)

resp
Ground Truth
resp
Prediction

Figure 3 (0.9832)

resp
Ground Truth
resp
Prediction

Figure 4 (0.4820)

resp
Ground Truth
resp
Prediction (Model Fails on point clouds without visibly defined edges)

Figure 5 (0.4467)

resp
Ground Truth
resp
Prediction (Model Fails on Assymetrical shapes, as evidenced due to the assymetry introduced due to the singular pillow on the right arm.)

Accuracy on Test Set: 91.13

3. Robustness Analysis

Classification: Rotating

New final test set accuracy: 0.7435 | Overall obsercations: Model is not rotation invariance

Figure 1

resp
Initial Prediction: Correct
resp
Rotated Ground Truth
resp
Prediction

Observation: Incorrect Prediction

Figure 2

resp
Initial Prediction: Correct
resp
Rotated Ground Truth
resp
Prediction

Observation: Incorrect Prediction

Figure 3

resp
Initial Prediction: Correct
resp
Rotated Ground Truth
resp
Prediction

Observation: Incorrect Prediction

Classification: 5000 Points

New final test set accuracy: 0.9698 | Overall observations: Model is able to adapt to varying point cloud densities, the model needs the general shape of the object in the point cloud to determine class label

Figure 1

resp
Initial Prediction: Correct
resp
Reduced Density

Observation: Incorrect Prediction

Figure 2

resp
Initial Prediction: Correct
resp
Reduced Density

Observation: Correct Prediction

Figure 3

resp
Initial Prediction: Correct
resp
Reduced Density

Observation: Correct Prediction

Figure 4

resp
Initial Prediction: Correct
resp
Reduced Density

Observation: Incorrect Prediction

Segmentation: 5000 Points

Figure 1 (0.9915 reduced to 0.9895)

resp
Ground Truth
resp
Initial Prediction
resp
Reduced Density Prediction

Figure 2 (0.9974 reduced to 0.9968)

resp
Ground Truth
resp
Prediction
resp
Reduced Density Prediction

Figure 3 (0.9832 increased marginally to 0.9941)

resp
Ground Truth
resp
Prediction
resp
Reduced Density Prediction

Figure 4 (0.4820 increased to 0.4924)

resp
Ground Truth
resp
Prediction (Model Fails on point clouds without visibly defined edges)
resp
Reduced Density Prediction (Model continues to fail on point clouds without visibly defined edges, reducing density further worsens this)

Figure 5 (0.4467 reduced to 0.4438)

resp
Ground Truth
resp
Prediction (Model Fails on Assymetrical shapes, as evidenced due to the assymetry introduced due to the singular pillow on the right arm.)
resp
Reduced Density Prediction (Model continues to fail on Assymetrical shapes, reduced point cloud density further worsens this.)

Reduced accuracy on Test Set: 0.9113 drops to 0.9042