NTU-2022Fall-DLCV

Deep Learning for Computer Vision 深度學習於電腦視覺 by Frank Wang 王鈺強

Surpassed strong baseline for all four assignments (Final grade: 97.22/100)

⭐Please consider starring this project if you find my code useful.⭐

Outline

For more details, refer to the reports.

HW1 Spec: Report
- Image classification ← Pretrained BEiT v1
  - Accuracy: 0.9360
- Image segmentation ← Pretrained Deeplab v3
  - mIoU: 0.7438
HW2 Spec: Report
- Face generation with GAN ← SNGAN (DCGAN with spectral normalization)
  - FID: 25.986
  - Face recognition accuracy: 0.9110
- Digit generation with diffusion model ← Ho et al. Classifier-Free Diffusion Guidance.
  - Digit classifier accuracy: 0.9990
- Domain adversarial network on MNIST-M, SVHN, and USPS
  - M→S Accuracy: 0.4943
  - M→U Accuracy: 0.9025
HW3 Spec: Report
- Zero-Shot image classification with CLIP ← CLIP L/14
  - Accuracy: 0.8124
- Image captioning with pretrained encoder ← Pretrained DeiT v3 as encoder
  - CIDEr score: 0.9413
  - CLIP score: 0.7310
- Attention map visualization for image captioning
HW4 Spec: Report
- 3D novel view synthesis ← DVGO (voxelized NeRF)
  - PSNR: 35.6029
  - SSIM: 0.9769
- Self-Supervised pretraining for image classification ← BYOL
  - Accuracy: 0.5985
  - Outperforms the supervised equivalent in both full fine-tuning and frozen backbone evaluation.
Final Project Spec -- Challenge 2: Poster
- Long-tailed 3D point cloud semantic segmentation