DLCVizsla

My NTU CommE 5052 Deep Learning for Computer Vision (by Prof. Frank Wang) homeworks,
you can get more details in each README.md or report.pdf inside folders.

The Vizsla (Hungarian: [ˈviʒlɒ]) is a dog breed originating in Hungary.

---

• demo_digit_recognition

  • MNIST handwritten digit recognition
  • for pytorch and cuda testing

• hw1_face_recognition

  • face images of 40 different subjects and 10 grayscale images for each subject, all of size (56, 46) pixels
  • cv2&matplotlib(gray), MSE, PCA, reconstruction, k-NN

• hw1_image_classification

  • 4 categories (classes) and 500 RGB images for each category, all of size (64, 64, 3) pixels
  • cv2&matplotlib(rgb), Bag of Word(BoW), Patches, k-means, PCA, scatterPlot3D, Soft-max(max pooling), k-NN

• hw2_YOLOv1_object_detection

  • https://github.com/dlcv-spring-2019/hw2-shannon112
  • YOLOv1: Redmon, Joseph, et al. "You only look once: Unified, real-time object detection." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
  • trained on DOTA-v1.5 Aerial Images
  • yolo loss, vgg16_bn+linear model

• hw3_dcgan_acgan_dann

  • https://github.com/dlcv-spring-2019/hw3-shannon112

  • GAN[1], DCGAN[2], ACGAN[3], DANN[4], GTA[5], tSNE plot

  • [1] Goodfellow, Ian, et al. "Generative adversarial nets." Advances in neural information processing systems. 2014.

  • [2] Radford, Alec, Luke Metz, and Soumith Chintala. "Unsupervised representation learning with deep convolutional generative adversarial networks." arXiv preprint arXiv:1511.06434 (2015).

  • [3] Odena, Augustus, Christopher Olah, and Jonathon Shlens. "Conditional image synthesis with auxiliary classifier gans." Proceedings of the 34th International Conference on Machine Learning-Volume 70. JMLR. org, 2017.

  • [4] Ganin, Yaroslav, et al. "Domain-adversarial training of neural networks." The Journal of Machine Learning Research 17.1 (2016): 2096-2030.

  • [5] Sankaranarayanan, Swami, et al. "Generate to adapt: Aligning domains using generative adversarial networks." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.

  • trained on USPS(28,28), MNIST-M(28,28,3), SVHN(28,28,3) Dataset

• hw4_rnn_action_recognition

  • https://github.com/dlcv-spring-2019/hw4-shannon112
  • pre-train resnet-50 + linear, LSTM, pack_padding, seq2seq action recognition, tSNE plot
  • train on 37 full-length videos (each 5-20 mins in 24 fps), and 4151 trimmed videos (each 5-20 secs in 24 fps), 11 action classes

• DLCV_Final

  • https://github.com/dlcv-spring-2019/final-ShiaGiBaLuanTrain
  • Cast Search by Portrait Challenge
  • face recognition (dlib), person re-id (resnet-50)