Instance Object Detection was developed using different techniques, such as:
- SIFT, ORB
- Template Matching
- Double-checking using Color Histograms, Zero Mean Normalized Cross Correlation (ZNCC), etc.
First, many different models were provided, such as:
and different noisy scenese, such as:
In the notebook available here, a study on the best filters for denoising and how to tune their parameters was conducted. Once the denoising was applied, this is how it looks:
At this point, an algorithm consisting of SIFT and template matching, customized to handle occluded objects, was used to filter out incorrect detections. Below is an example of an occluded object that has been correctly detected:
The same approach as for single instance detection was applied, but for multiple instance object detection. This time, the algorithm is slightly more complex than the one in the first part, utilizing additional criteria for double-checking SIFT detections, comparing color distributions, and using similarity functions. An example is shown below:
Implementation of a neural network that classifies smartphone pictures of products found in grocery stores. It's divided into two parts: first, implementation from scratch of my own neural network for image classification; then, you fine-tuning of a pretrained network provided by PyTorch.
A VGG-like architecture was developed, and preprocessing was applied to the data. Techniques such as Dropout Regularization, Global Average Pooling, Batch Normalization, an Adam learning rate scheduler, and LeakyReLU as the activation function were utilized to achieve higher accuracy. We achieved an accuracy of approximately 67.5% on the test set.
Loss and Accuracy over Validation Set during training:
A pre-trained ResNet-18 model on ImageNet-1k (V1) was used and fine-tuned. By adjusting fine-tuning parameters, changing the image resolution, applying data augmentation and weight decay, and experimenting with freezing layers, we managed to achieve approximately 88% accuracy on the Test Set!
Loss and Accuracy over Validation Set during training:
Example of classification:
