1 Introduction
1.1 Theories of visual perception
1.2 Related work
1.3 Contributions
2 Proposed approach
2.1 Gestalt Interest Points Detection
2.2 Gestalt Interest Points Description
2.2.1 Inter-GIP distances (IGD)
2.2.2 Gestalt regions of interest (GROI)
3 Experiments and results
3.1 Overview of experiments
3.2 GIP for weight-invariant face recognition
3.2.1 Dataset
3.2.2 Experimental setup
3.2.3 Evaluation
Method | Acc. (%) | Accuracy 30\(^\circ\) rotated (%) | Average number of description values per face |
---|---|---|---|
BoVW + SIFT | 20 | 13.3 | 25,309 |
BoVW + SURF | 33 | 13.3 | 57,984 |
BoVW + MSER | 6.7 | 6.7 | 132 |
BoVW + FREAK | 20 | 20 | 59,473 |
BoVW + ORB | 13.3 | 13.3 | 8883 |
BoVW + GIP | 53.3 | 53.3 | 52,536 |
BoVW + GIP \(t=70\) | 53.3 | 53.3 | 23,086 |
BoVW + GIP \(\alpha =0.0009\) | 46.7 | 46.7 | 23,101 |
BoVW + GIP \(t=70\) \(\alpha =0.0009\) | 46.7 | 46.7 | 10,425 |
3.3 GIP-IGD for image categorization
3.3.1 Datasets
3.3.2 Experimental setup
3.3.3 Evaluation
3.4 Deep Gestalt regions of interest for makeup-robust face recognition
3.4.1 Dataset
3.4.2 Experimental setup
Layer name (type) | Output shape |
---|---|
conv2d_1 (Conv2D) | (158, 158, 32) |
conv2d_2 (Conv2D) | (156, 156, 32) |
max_pooling2d_1 (MaxPooling2) | (78, 78, 32) |
dropout_1 (Dropout) | (78, 78, 32) |
conv2d_3 (Conv2D) | (76, 76, 64) |
conv2d_4 (Conv2D) | (74, 74, 64) |
max_pooling2d_2 (MaxPooling2) | (37, 37, 64) |
dropout_2 (Dropout) | (37, 37, 64) |
conv2d_5 (Conv2D) | (35, 35, 64) |
max_pooling2d_3 (MaxPooling2) | (35, 11, 64) |
dropout_3 (Dropout) | (35, 11, 64) |
conv2d_6 (Conv2D) | (35, 10, 64) |
max_pooling2d_4 (MaxPooling2) | (17, 10, 64) |
dropout_4 (Dropout) | (17, 10, 64) |
flatten_1 (Flatten) | (10880) |
dense_1 (Dense) | (256) |
dropout_5 (Dropout) | (256) |
dense_2 (Dense) | (6) |