search

kris-singh / Deep-Unsupervised-Saliency-Detection

36 stars 8 forks source link

readme

[WIP] Implmentation of the Deep Unsupervised Saliency Detection: A Multiple Noisy Labeling Perspective

Requirements

  1. pytorch 1.0
  2. python 3.6

  3. numpy Dataset

  4. MSRA-B

Results

Table of Contents

  1. Training on Small Dataset(10 Images)
  2. Training on Large Dataset(1500 Images Training, 500 Images Validataion, 500 Images Test)

Training on Small Dataset(10 Images)

Experiment Performed for overfitting, checking if the model works, tested for all models, Only reporting for the full(Training Noise module)

Training on Large Dataset(1500 Images Training, 500 Images Validataion, 500 Images Test)

The most important part here is the scheduler, since we keep on training with the same leanring rate without, Image size is taken to be 256, Number of Epochs is 20, Early Stopping on validation loss, paitence of 5 epochs

  1. Exp-1 Using Adam Optimizer

    Exp Name Optimizer Batch Size LR Betas Momentum Scheduler Notes Decay Factor Paitence THRESHOLD MINLR COOLDOWN Recall-Test Precision-Test F1-Test MAE-TEST
    Real Adam 16 3e-4 (0.9, 0.99) x ReduceLROnPlateu Label is the ground truth 0.9 1 1e-4 1e-16 1 0.5023 0.9622 0.66 0.035
    Noise Adam 8 3e-4 (0.9, 0.99) x ReduceLROnPlateu Use all Noise labesl 0.9 1 1e-4 1e-16 1 0.793 0.946 0.862 0.028
    Avg Adam 16 3e-4 (0.9, 0.99) x ReduceLROnPlateu Use Avg of Noise Labels 0.9 1 1e-4 1e-16 1 0.802 0.907 0.851 0.041
    Full Adam 4 3e-4 (0.9, 0.99) x ReduceLROnPlateu Full Training 0.9 1 1e-4 1e-16 1 0.841 0.857 0.848 0.036

    Sample Images img Average Exp Sample Maps img Average Exp Sample Threshold@0.5 img Real Exp Sample Maps img Real Exp Sample Maps Threshold@0.5 img Noise Exp Sample Maps img Noise Exp Sample Maps Thresold@0.5 img Full Exp Sample Maps img Full Exp Sample Maps Thresold@0.5 img

  2. Exp-2 Using SGD Optimizer

    Exp Name Optimizer Batch Size LR Betas Momentum Scheduler Notes Decay Factor Paitence THRESHOLD MINLR COOLDOWN Recall-Test Precision-Test F1-Test MAE-TEST
    Real SGD 16 1e-3 x 0.9 ReduceLROnPlateu Label is the ground truth 0.9 1 1e-4 1e-16 1 0.865 0.907 0.885 0.032
    Noise SGD 8 1e-3 x 0.9 ReduceLROnPlateu Use all Noise labesl 0.9 1 1e-4 1e-16 1 0.620 0.820 0.706 0.044
    Avg SGD 16 1e-3 x 0.9 ReduceLROnPlateu Use Avg of Noise Labels 0.9 1 1e-4 1e-16 1 0.934 0.639 0.758 0.058
    Full SGD 4 1e-3 x 0.9 ReduceLROnPlateu Full Training 0.9 1 1e-4 1e-16 1 0.825 0.795 0.809 0.027

Average Exp Sample Maps Average Exp Sample Maps Average Exp Sample Threshold@0.5 Average Exp Sample Threshold 0.5 Real Exp Sample Maps Real Exp Sample Maps Real Exp Sample Maps Threshold@0.5 Real Exp Sample Maps Threshold 0.5 Noise Exp Sample Maps Noise Exp Sample Maps Noise Exp Sample Maps Thresold@0.5 Noise Exp Sample Maps Thresold 0.5 Full Exp Sample Maps Full Exp Sample Maps Full Exp Sample Maps Thresold@0.5 Full Exp Sample Maps Thresold 0.5