Synthetic Data

The image below shows sample images generated by rendering a CAD model using PyTorch3D. The first column is the rendered RGB image. The second column is the corresponding depth map. The third column shows the corresponding heatmaps indicating the 2D locations of visible keypoints.

Examples of data generated

Qualitative Analysis

The image below shows the keypoints predicted for a wide range of aircraft.

Predicted keypoints on images in validation dataset (in red)

Quantitative Analysis

The table below shows the Percentage of Correct Keypoints (PCK) and Mean Euclidean Distance(MAD). PCK measures the percentage of keypoint predictions that lie within a threshold distance of the ground-truth keypoints. MAD measures the average distance between locations of keypoint predictions and their corresponding ground truth. All the metrics are computed on validation split in PASCAL3D+ dataset for the images belonging to the aeroplane class.

PCK_d=3	PCK_d=5	PCK_d=10	MAD
46.4	57.4	63.5	31.4

Metrics on the PASCAL3D+ dataset