|
|
Counting Crates in Images
Mičulek, Petr ; Špaňhel, Jakub (oponent) ; Herout, Adam (vedoucí práce)
This thesis deals with the topic of using deep learning to count crates in images. I have designed a crate-counting solution for blocks of matchboxes, using a fully convolutional classification-based network with a high resolution output. The original project proposition counted on using a dataset of photos of crates from a beer brewery warehouse. I did not get access to the dataset in the end. On the recommendation of my supervisor, I based the crate-counting solution on a custom dataset of matchbox photos. The CNN is trained using image patches, leading to a fast solution working even on smaller datasets. Matchbox keypoints are detected by the CNN in the input images and they are processed by a keypoint estimation and crate-counting algorithm to produce the final crate count. On validation data, the solution has a 12.5% failure rate and a MAE of 11.14. Thorough experimentation was performed to evaluate the solution and the results verify that this approach can be used for object counting.
|
|
|
Explainable Face Liveness Classification
Mičulek, Petr ; Beran, Vítězslav (oponent) ; Špaňhel, Jakub (vedoucí práce)
The goal of this thesis is to explore, develop, and evaluate explainable face presentation attack detection (PAD) systems. PAD systems act as security filters for face recognition, preventing spoofed faces from reaching the identification phase. These systems are a necessary component enabling the recent rise of biometric systems used in smartphones and security cameras. While neural networks are the standard method for this task, they are commonly a black-box method providing no explanation. To provide a better understanding of the detection process, input attribution methods are applied. Their suitability is studied and various variants are compared. Of the seven methods compared, GradCAM using test-time augmentation is evaluated as the best, achieving a deletion metric AUC of 0.658 and an insertion metric AUC of 0.908. Experiments with the explanations show their limited capability at helping understand the model, but provide hints at how the predictive accuracy of the PAD system can be verified, and possibly improved.
|
|
|
Counting Crates in Images
Mičulek, Petr ; Špaňhel, Jakub (oponent) ; Herout, Adam (vedoucí práce)
This thesis deals with the topic of using deep learning to count crates in images. I have designed a crate-counting solution for blocks of matchboxes, using a fully convolutional classification-based network with a high resolution output. The original project proposition counted on using a dataset of photos of crates from a beer brewery warehouse. I did not get access to the dataset in the end. On the recommendation of my supervisor, I based the crate-counting solution on a custom dataset of matchbox photos. The CNN is trained using image patches, leading to a fast solution working even on smaller datasets. Matchbox keypoints are detected by the CNN in the input images and they are processed by a keypoint estimation and crate-counting algorithm to produce the final crate count. On validation data, the solution has a 12.5% failure rate and a MAE of 11.14. Thorough experimentation was performed to evaluate the solution and the results verify that this approach can be used for object counting.
|
host ::
RSS.