2023 Spring NTU CSIE

• Objective: Implement object detection using the DINO framework.
• Techniques and Implementation:
  • Employed advanced object detection methods using the DINO (DEtection with Transformers) framework.
  • Re-trained the model on a given dataset to tailor it for specific object detection tasks.
  • Utilized a pretrained model, checkpoint0033_4scale.pth, fine-tuned on the COCO 2017 dataset.
  • Backbone architecture used was R50.
• Performance Metrics:
  • Achieved a mean Average Precision (mAP) of 0.5233.

• Objective: Analyzed convergence issues in DETR models and explored advanced models and techniques.
• Key Highlights:
  • Analyzed DETR (End-to-End Object Detection with Transformers) for convergence issues.
  • Compared DAB-DETR and DN-DETR, addressing slow convergence of DETR.
  • Explored knowledge distillation methods, both from logits and intermediate layers.
  • Discussed CLIP's training process and zero-shot classification applications.
  • Investigated challenges in Open-Vocabulary Object Detection with RegionCLIP.

• Objective: Addressed data imbalance in object detection using foundation models and data augmentation techniques.
• Key Features:
  • Utilized BLIP2 for image captioning and GLIGEN for data augmentation.
  • Addressed data imbalance from HW1's dataset using generated prompts for image generation.
  • Applied text-to-image generation techniques to augment object detection datasets.
  • Evaluated using Fréchet Inception Distance (FID) for the quality of generated images.
  • Improved detection model performance post data augmentation.
  • Demonstrated techniques through examples of image captioning and text-to-image generation.