Homework-5. Tianyuan Zhang

andrew id: tianyuaz

Problem-1

I used a PointNet-like model with 3 Linear layer and one global average pooling to get the abstract global features, and use that for classification

Final test accuracy: 92.76% at epoch 200

Showcase:

GT: Chair; Prediction: Chair

Showcase faliure:

GT: Chair; Prediction: Lamp

Interepreation: Maybe the chair is tooooo tall

GT: Vases; Prediction: Lamp

Interepretation: Maybe the leaf of the flower looks like the large lamp head..

GT: Lamp; Prediction: Vases

Interepretation: I don't know why this is wrong, I didn't think this has any similarity as Vases..

Problem-2

I fuse the global average pooling features with the per-point feature, and add more linear layers to get a per-point classification prediction.

Final test accuray: 80.02% at epoch 200

Showcase:

Easy examples:

GT ------ Prediction; Accuracy: 88.49%

GT ------ Prediction; Accuracy: 92.57%

GT ------ Prediction; Accuracy: 80.2%

Interepreation: These three are just standarized easy chairs. And the size is quites similar, maybe this is why it is easy.

More on hard examples:

GT ------ Prediction; Accuracy: 63.16%

Interepretation: It cannot distinguish the lower part of the chair.

GT ------ Prediction; Accuracy: 58.37%

Interepretation: It also cannot distinguish the lower part of the chair.

Problem-3

Test with small amount of rotations

I try to rotate the point clouds along z, y, x axis for 15 degree and test the accuracy:

I implement them in eval_seg_rotate.py and eval_rotate_cls.py

For clssification models, accuracy drops from 92.76% to 81.53 %

For segmentation model, accuracy drops from 80.02% to 71.72 %

Sampling fewer points per scene when test

When changing the number of sampled points per instance from 10000 to 1000 For clssification models, accuracy drops from 92.76% to 92.23 %

For segmentation model, accuracy drops from 80.02% to 79.99 %

When changing the number of sampled points per instance from 10000 to 200 For clssification models, accuracy drops from 92.76% to 91.61 %

For segmentation model, accuracy drops from 80.02% to 79.86%

seems the molde is quite robust to the number of points. Or say, hundereds of points is enough to recognize the objects!