omkarchittar/PointCloud_Classification_and_Segmentation

Implementing a PointNet based architecture for classification and segmentation with point clouds. Q1 and Q2 focus on implementing, training and testing models. Q3 asks you to quantitatively analyze model robustness.

Point Cloud Classification and Segmentation

Name: Omkar Chittar

PointCloud_Classification_and_Segmentation
+-checkpoints
+-data
+-logs
+-output
+-output_cls
+-output_cls_numpoints
+-output_cls_rotated
+-output_seg
+-output_seg_numpoints
+-output_seg_rotated
+-README.md
+-report
data_loader.py
eval_cls_numpoints.py
eval_cls_rotated.py
eval_cls.py
eval_seg_numpoints.yml
eval_seg_rotated.py
eval_seg.py
models.py
train.py
utils.py

Installation

• Download and extract the files.
• Make sure you meet all the requirements given on: https://github.com/848f-3DVision/assignment2/tree/main
• Or reinstall the necessary stuff using 'environment.yml':

conda env create -f environment.yml
conda activate pytorch3d-env

Data Preparation

Download zip file (~2GB) from https://drive.google.com/file/d/1wXOgwM_rrEYJfelzuuCkRfMmR0J7vLq_/view?usp=sharing. Put the unzipped data folder under root directory. There are two folders (cls and seg) corresponding to two tasks, each of which contains .npy files for training and testing.

• The data folder consists of all the data necessary for the code.
• There are 6 output folders:
  1. output_cls folder has all the images/gifs generated after running eval_cls.py.
  2. output_seg folder has all the images/gifs generated after running eval_seg.py.
  3. output_cls_numpoints folder has all the images/gifs generated after running eval_cls_numpoints.py.
  4. output_seg_numpoints folder has all the images/gifs generated after running eval_seg_numpoints.py.
  5. output_cls_rotated folder has all the images/gifs generated after running eval_cls_rotated.py.
  6. output_seg_rotated folder has all the images/gifs generated after running eval_seg_rotated.py.
• All the necessary instructions for running the code are given in README.md.
• The folder report has the html file that leads to the webpage.

1. Classification Model

I implemented the PointNet architecture.

• After making changes to:
  1. models.py
  2. train.py
  3. eval_cls.py

Run the code:

python train.py --task cls

The code trains the model for the classification task.

Evaluate the trained model by running the code:

python eval_cls.py

Evaluates the model for the classification task by rendering point clouds named with their ground truth class and their respective predicted class. Displays the accuracy of the trained model in the terminal. The rendered point clouds are saved in the output_cls folder.

The test accuracy of the model is stored as best_model.pt in the ./checkpoints/cls folder and has a value 0.9769

Results

Correct Classifications

Point Cloud			Ground truth Class	Predicted Class
			Chair	Chair
			Vase	Vase
			Lamp	Lamp

Incorrect Classifications

Point Cloud	Ground Truth Class	Predicted Class
	Chair	Lamp
	Vase	Lamp
	Lamp	Vase

Analysis

The misclassifications made by the PointNet model on the few failure cases seem to be due to those examples deviating significantly from the norm for their respective categories. For instance, the misclassified chair examples have unusual or atypical designs - one is folded up and missing a seat, while the other is unusually tall. Similarly, some of the misclassified vases and lamps have shapes that overlap more with the opposing class. Additionally, the chair class appears to have less shape variety overall compared to vases and lamps. Chairs components tend to be more standardized (seat, legs, back, etc). In contrast, the vase and lamp categories exhibit greater diversity in proportions and silhouettes (floor lamps vs desk lamps, vases with or without flowers etc). The model's confusion between these two classes likely stems from their greater morphological similarity in many cases - symmetry about the vertical axis, cylindrical profiles etc.

2. Segmentation Model

I implemented the PointNet architecture.

• After making changes to:
  1. models.py
  2. train.py
  3. eval_seg.py

Run the code:

python train.py --task seg

The code trains the model for the Segmentation task.

Evaluate the trained model by running the code:

python eval_seg.py

Evaluates the model for the Segmentation task by rendering point clouds with segmented areas with different colors. The rendered point clouds are saved in the output_seg folder. Displays the accuracy of the trained model in the terminal.

The test accuracy of the model is stored as best_model.pt in the ./checkpoints/seg folder and has a value 0.9022

Results

Good Predictions

Ground truth point cloud	Predicted point cloud	Accuracy
		0.9836
		0.9237
		0.917

Bad Predictions

Ground truth point cloud	Predicted point cloud	Accuracy
		0.5171
		0.4776
		0.5126

Analysis:

The model struggles to accurately segment sofa-like chairs where the boundaries between components like the back, headrest, armrests, seat and legs are less defined. The blending of these parts without clear delineation poses a challenge. Similarly, chairs with highly irregular or atypical shapes and geometries also confuse the model as they deviate significantly from the distribution of point clouds seen during training. On the other hand, the model performs very well in segmenting chairs with distinct, well-separated components like a distinct back, seat, separable arm rests and discrete legs. Chairs that have intricate details or accessories that overlap multiple segments, like a pillow over the seat and back, trip up the model. In such cases, there is often bleeding between segments, with the model unable to constrain a larger segment from encroaching on adjacent smaller segments.

3. Robustness Analysis

3.1. Rotating the point clouds

Here we try to evaluate the accuracy of the classification as well as the segmentation models by rotating the point clouds around any one axis (x/y/z) or their permutations. Rotate each evaluation point cloud around x-axis for 30, 60 and 90 degrees. I have written the code to loop over specific object indices and while looping over various thetas (angles).

3.1.1. Classification

Run the code:

python eval_cls_rotated.py

Evaluates the model with rotated inputs for the classification task by rendering point clouds named with their rotated angle, ground truth class and their respective predicted class. Displays the accuracy of the trained model in the terminal. The rendered point clouds are saved in the output_cls_rotated folder.

Class	Ground Truth	30 deg	60 deg	90 deg
Chair
Vase
Lamp
Test Accuracy	0.9769	0.7992	0.2235	0.3012

3.1.2. Segmentation

Run the code:

python eval_seg_rotated.py

Evaluates the model with rotated inputs for the segmentation task by rendering point clouds named with their rotated angle, and prediction accuracy. Displays the accuracy of the trained model in the terminal. The rendered point clouds are saved in the output_seg_rotated folder.

	0 deg	30 deg	60 deg	90 deg
Test Accuracy	0.9022	0.7992	0.399	0.1319

Analysis

The model struggles to make accurate predictions when the point cloud is rotated dramatically away from an upright orientation. This limitation is likely due to the lack of data augmentation during training to include non-upright point cloud configurations. Without exposure to rotated variants of the object classes, the model fails to generalize to point clouds that deviate hugely from the expected upright positioning seen in the training data. Incorporating point cloud rotations during training data generation would likely improve the model's ability to recognize and segment objects despite major shifts in orientation. By augmenting the data to simulate tilted, skewed or even completely inverted point clouds, the model could become invariant to orientation and handle such cases gracefully during prediction.

3.2. Varying the sampled points in the point clouds

Here we try to evaluate the accuracy of the classification as well as the segmentation models by varying the number of sampled points in the point clouds. I have written the code to loop over specific object indices and while looping over various num_points.

3.2.1. Classification

Run the code:

python eval_cls_numpoints.py

Evaluates the model with varying number of sampled points inputs for the classification task by rendering point clouds named with their index, number of points, ground truth class and their respective predicted class. Displays the accuracy of the trained model in the terminal. The rendered point clouds are saved in the output_cls_numpoints folder.

Class	10	100	1000	10000
Chair
Vase
Lamp
Test Accuracy	0.5012	0.8255	0.8992	0.9769

3.2.2. Segmentation

Run the code:

python eval_seg_numpoints.py

Evaluates the model with varying number of sampled points inputs for the segmentation task by rendering point clouds named with their index, number of points, and their respective predicted class accuracy. Displays the accuracy of the trained model in the terminal. The rendered point clouds are saved in the output_seg_numpoints folder.

	10	100	1000	10000
Test Accuracy	0.4673	0.7992	0.8599	0.9022

Analysis

The model demonstrates considerable robustness to sparsity in the point cloud inputs. With as few as 10 points, it can achieve 25% test accuracy, rising rapidly to 80% accuracy with only 50 points. This suggests the model is able to infer the correct shape from even a very sparse sampling of points. However, its ability to generalize from such limited information may be challenged as the number of classes increases. Discriminating between more categories with fewer representative points could lower the accuracy, despite the architectural design choices to promote invariance to input sparsity. There may be a threshold minimum density below which the lack of key shape features impedes reliable classification, especially with additional object classes added. But with this 3 class dataset, the model's performance from merely 50 input points indicates surprising generalizability from scant point data.

4. Webpage

The html code for the webpage is stored in the report folder along with the images/gifs. Clicking on the webpage.md.html file will take you directly to the webpage.