Object landmark discovery through unsupervised adaptation

Authors: Enrique Sanchez, Georgios Tzimiropoulos

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This section describes the experimental set-up carried out to validate the proposed approach (Sec. 4.1), as well as the obtained results (Sec. 4.2). We show that our method surpasses fully unsupervised techniques trained from scratch as well as a strong baseline based on fine-tuning, and produces state-of-the-art results on several datasets.
Researcher Affiliation Collaboration 1 Samsung AI Centre Cambridge, UK {e.lozano, georgios.t}@samsung.com Georgios Tzimiropoulos1,2 2 Computer Vision Lab University of Nottingham, UK yorgos.tzimiropoulos@nottingham.ac.uk
Pseudocode No The paper describes its methods verbally and mathematically but does not include any pseudocode or algorithm blocks.
Open Source Code Yes Code can be found at tiny.cc/Git Hub-Unsupervised.
Open Datasets Yes For training the object landmark detectors in an unsupervised way, we used the Celeb A [18], the UT-Zappos50k [41, 40], and the Cats Head datasets [42]. ... For our method, the landmark detector is firstly pre-trained on the task of Human Pose Estimation. In particular, the network is trained to detect K = 16 keypoints, corresponding to the human body joints, on the MPII training set [2].
Dataset Splits Yes For the UT-Zappos50k, we used 49.5k and 500 images to train and test, respectively [35, 43]. Finally, for the Cats Head dataset, we used four subfolders to train the network ( 6, 250 images), and three to test it (3, 750 images). For MAFL, we used the official train/test partitions. For AFLW, we used the same partitions as in [13]. For LS3D, we used the partitions as defined in [3].
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, memory, or processor types used for running experiments.
Software Dependencies No All networks are implemented in Py Torch [23]. (No version number for PyTorch or other libraries is provided.)
Experiment Setup Yes Training: We generate the pairs (y, y ) by applying random similarity transformations (scaling, rotation, translation) to the input image. We used the Adam optimizer [16], with (β1, β2) = (0, 0.9), and a batch size of 48 samples. The model is trained for 80 epochs, each consisting of 2, 500 iterations, with a learning rate decay of 0.1 every 30 epochs. ... The output spatial resolution is 32 32, which is converted into a K 2 matrix of coordinates with a softargmax layer (β = 10). The coordinates are mapped back to heatmaps using σ = 0.5. In all of our experiments, K is set to 10 points.