Skip to main content

How to improve Object Detection performance? + tools

· 3 min read
SeulGi Hong

Episode 1 for my VIPrior challenge journey link!

Useful Papers​

  1. Making Images Real Again: A Comprehensive Survey on Deep Image Composition
  2. Bag of Freebies for Training Object Detection Neural Networks

Image Composition?​

  • The issue of small datasets -> The concept of image composition came up as a way to diversify information through data augmentation.
  • (Refer to Paper 1) It was shared by a team member that image compositing is not about randomly placing objects like simple copy & paste algorithm. Instead, it aims to make the composition more realistic.
    • For example, human should be not copy&pasted in the sky, since we cannot fly!

Tools​

  • I came across a website called FiftyOne. It seems to be related to detection bounding boxes and instance segmentation annotation tools. I should take a look at it.
  • Roboflow: It provides detailed analysis of the dataset itself, various augmentations, and the ability to train models on the web.
    • The visualization is also well-done, making it convenient.
    • It seems that training is free for up to 3 times, and an email will be sent when the training is completed.
    • I haven't used either of them yet, so I'll make a note after trying them out.
      • -> (UPDATED) I totally fall in love to the Roboflow!

Hyperparameter Tuning?​

  • Refer to Paper 2.
  • (My teammate who keep on this field told me that) It is recommended to use well-known hyperparameters as they are, instead of hyperparameter tuning and search.

Small Object, High Bounding Box Overlap Issue​

When training for small objects, a simple method is to increase the resolution. We had a meeting to analyze the characteristics of the current target data (VIPriors link, for challenge) through the dataset health check in Roboflow and find possible solutions.

The discussion content, future improvements, and work direction will be added below after the meeting.

Idea? (after meeting)​

  • Train based on their own characteristics: Divide into small, medium, and large categories.
    • for example, in the case of bicycle data where there are many overlapping small object bounding boxes, divide them into multiple models, each responsible for a few classes. Then, combine the results during inference.
    • It is not an ensemble, simply divide the recognizable classes by several models.
  • Transfer learning: Pretrained models are not allowed in this challenge. But, instead of random initialization, what if we crop object bounding boxes and train an image classification model to initialize the image encoder?

etc​

survey on crowd detection...