UCL Computer Science


MEng Computer Science Student project work published at BMVC 2021

26 October 2021

Congratulations to undergraduate student Seunghoi Kim, whose paper has been accepted to the 2021 BMVC as a first author!

Suenghoi Kim

The British Machine Vision Conference (BMVC) is the British Machine Vision Association (BMVA) annual conference on machine vision, image processing, and pattern recognition. It is one of the major international conferences on computer vision and related areas held in the UK. With increasing popularity and quality, it has established itself as a prestigious event on the vision calendar.

BMVC logo

Seunghoi explained - “This paper is an extension of my undergraduate thesis, ‘GAGCN: Graph Convolutional Network for 3D Point Cloud Segmentation’, which was awarded ‘Outstanding Paper’ Prize sponsored by IBM & UCL in July 2021. I developed this into the current paper, ‘AGCN: Adversarial Graph Convolutional Network for 3D Point Cloud Segmentation’ which has recently been accepted to British Machine Vision Conference (BMVC) 2021 as a first author, holding in November."

Below is the summary of Suenghoi’s paper:

3D Point clouds are one of the most popular 3D representations as they can preserve the original geometric representation with minimal information loss and object segmentation provides a high level understanding of object structure valuable in various applications such as medicine, robotics and self-driving car.

However, many current 3D point cloud networks pose problems such as low segmentation accuracy or high complexities due to their crude network architectures, predicting labels independently and local feature aggregation method, losing an important local geometric information.

To overcome this, we propose:

  1. A graph convolutional network in an adversarial learning scheme where a discriminator network provides a segmentation network with informative information to improve neighbourhood labelling consistency.
  2. A graph convolution, GeoEdgeConv, as a means of local feature aggregation method to improve segmentation accuracy and model complexity.
  3. A lightweight version, AGCN-S, which achieves the smallest model complexity compared to the SOTA by using efficient operations such as Density-aware Random Sampling, Grouped and Dilated convolution.

We use an embedding L2 loss as an adversarial loss and by learning to reduce the high level feature difference between the ground truth and the predicted output, the predictions are less noisy and has better neighbourhood label consistency. Our GeoEdgeConv preserves geometric structures over convolution layers by using both point and relative position features, which helps to learn fine details of complex structures, and thus improves segmentation accuracy in boundaries and reduces label noise inside a class without increased computational complexity.  

In our experiment, our model outperforms the state-of-the-art with lower complexity and shows strong prospects in applications requiring low power but high segmentation performance.

Seunghoi’s supervisor Professor Daniel Alexander commented:

“I am delighted that Seunghoi’s hard work has been rewarded with getting his MEng project work published at BMVC. This is an outstanding achievement that researchers several years into a PhD would be proud of. Seunghoi has a great scientific career ahead of him.”

Professor Daniel Alexander – Project Supervisor / Director, UCL CMIC

Overall architecture of the proposed model


Proposed local feature extraction method for 3D point cloud