Questions tagged [computer-vision]

Questions related to image representation, segmentation, visual object categorization and image processing algorithms in general.

Filter by
Sorted by
Tagged with
0 votes
0 answers
19 views

α-balanced focal loss - why we actually decrease the importance of positive class

This is the equation for Focal Loss. The loss is an extension of weighted cross entropy, and aims to balance the impact of majority of easy negative class samples. The α parameter is a weighing term ...
  • 626
0 votes
0 answers
10 views

Calculate uncertainty of surface normal based on points uncertainty

How to calculate angular variation of fitted plane from points that has positional uncertainty with normal probability distribution? For example, I have 4 points P1 = 102.0000 84.0000 139.5443 P2 = ...
0 votes
0 answers
6 views

training on grey scale images and perform detection on RGB images

Could be a problem training an algorithm such as YOLO to perform object detection on grey scale images and then apply it on RGB images?
0 votes
0 answers
5 views

What is 6DoF pose estimation and what's the difference between 6DoF pose estimation and 3D object detection?

For objection dectection, we need to provide the bounding boxes of given object, plus confide score and category. And for 3d object detection, the bounding cubes and rotations are provided. As the ...
  • 111
2 votes
0 answers
15 views

How can fully convolution neural networks handle images of different sizes?

I've read that if we want to use images of different sizes in a convolutional neural network without resizing the images to a default size, we can use Fully Convolutional Neural Networks. But I do not ...
  • 121
0 votes
0 answers
21 views

Can the Dice coefficient be used for Mean Average Precision for Instance Segmentation?

I'm a beginner to computer vision currently working out a 2D multi-class instance segmentation problem on an imbalanced dataset of images with masks (98% background, 6 object classes for the remaining ...
0 votes
0 answers
4 views

About image super resolution task categories

What is the difference between classical SR, lightweight SR, real-world SR? As I know, these tasks use a low-resolution image as an input and get a high-resolution image as an output. Papers use even ...
  • 11
0 votes
0 answers
7 views

Pseudo Label Generation for Unsupervised Video Anomaly Detection

I am trying to implement this paper for unsupervised video anomaly detection. The gist of the paper seems to be: Create a dataset for an unsupervised setting, by mixing up the train and anomalous ...
  • 111
1 vote
1 answer
16 views

Use the Same Learning Rate to Train All Models When Doing Experiments For A Deep Learning Paper?

When writing a deep learning paper, I need to train several CNN models and compare their performances. They are from different architectures so different designs. I'm wondering should I use the same ...
  • 111
0 votes
0 answers
12 views

Image preprocessing guidance for a Multi-class image classification problem

Scenario Multi-class image classification of mechanical parts. The Train set contains images of parts on a white background. The Test set contains images of parts from the workshop. Parts are ...
  • 111
1 vote
0 answers
78 views

What is the correct way of normalizing/standardizing image-like data?

I have image-like data (e.g. H x W x C), where each channel contains quite different information. You can think of it being a 2D map (H x W) with information like elevation, wind velocity, temperature ...
0 votes
1 answer
30 views

plotting the latent space of a GAN

I am working on gans and wanted to know how I can plot the latent space of gan. Like I have a latent space of shape (50,250). So it is an n-d array of length 50 and 250 points representing each one of ...
  • 1
0 votes
0 answers
16 views

Can I deal with datasets of images with different sizes with pre-trained densenet?

I have a dataset of images with very different sizes, but that are bigger than 224 X 224. It seems that the pre-trained densenet model of pytorch can accept images of different sizes during training ...
  • 566
1 vote
0 answers
34 views

How ResNet researchers achieve high accuracy?

I tried to recreate the original ResNet 50 model myself, using all the same layers, hyperparameters, and data augmentation methods mentioned in the paper. (I only trained on a few hundred iterations ...
0 votes
0 answers
8 views

Does the GTSRB dataset duplicate images?

I'm confused on the existence of duplicate images (i.e., images of the same physical traffic sign) in the GTSRB dataset. Their website says: Physical traffic sign instances are unique within the ...
  • 153
0 votes
0 answers
8 views

Using optical flow to predict velocities

I am no expert in this field but more of a beginner with a bit of experience, so please keep the answer as simple as possible. I cannot be very specific about this topic but what I am trying to do is ...
0 votes
1 answer
24 views

Can bounding boxes be used in UNets?

I recently read a paper where the researchers used a UNet algorithm to localize/detect cyclones using a bounding box. However, my interpretation of a UNet is that it performs semantic segmentation and ...
  • 1
0 votes
1 answer
31 views

Which tool is more suitable for visualizing the distribution of multiple real and synthetic image datasets, t-SNE or PCA?

I am doing a thesis on the generation of synthetic data for training a deep learning model and evaluating it on real data. I have a few different real datasets, and I generated multiple synthetic ...
  • 147
1 vote
0 answers
31 views

Yolov3 targets build, Do all 3 scales need to have an anchor box designed for each object in the image or just the scale with better IoU? [closed]

I'm implementing Yolo v3 in Pytorch, Imagine we have an image with a large object in it, whem building targets, I'll have to assign the anchor box with highest IoU with this object for each scale, (So ...
  • 11
0 votes
0 answers
16 views

Data split when using pseudo-labeling semi-supervised learning method

I'm trying to train a 3D segmentation model. The dataset I own consists of small number of labeled samples(~21) and a lot of unlabeled samples(~200). I'm using a simple semi-supervised method, where I'...
1 vote
0 answers
28 views

Calculating output from kernel

Given the following task: You feed an image $I$ with the dimensions $200\text{x}200\text{x}3$ in the convolutional layer, which consists of $12323$ kernels $K$ each with the size $40\text{x}40\text{x}...
  • 11
1 vote
0 answers
14 views

"Biasing" a generalized trained object detector towards specific examples during inference

I have trained an object detection deep learning model on many different types of cars (shape, color, car model variations etc.). I'm just using a single class "Car" for all the different ...
  • 11
1 vote
1 answer
62 views

Splitting medical dataset by patient

I am currently trying to train a CNN model to classify CT-scans. I split the dataset using K-fold cross-validation and since the dataset I am using contains multiple slices per patient, I split the ...
1 vote
0 answers
30 views

Autoencoder for Image Classification/Clustering

I'm trying to construct a convolutional autoencoder that will learn features of an image dataset in the low-dimensional latent space. My hope is to use the latent vectors to cluster the images into ...
  • 111
2 votes
2 answers
89 views

few images in validation and test set

I've a dataset with about 123 images (two categories, 19 defect and 104 no defect). I've to implement a classifier so I've decided to split my data in train (70% of all data), validation (20% of all ...
0 votes
1 answer
58 views

What is the purpose of 2 fully connected hidden layers in VGG16?

Here is the architecture of VGG16: The first 18 convolution layers can be understood as feature extraction. How about the two fully connect hidden layers after them? What is their purpose?
  • 867
0 votes
0 answers
27 views

How to perform image classification in a dataset of images with heterogeneous sizes?

I have a dataset of images with very different sizes (ranging from 100X100 pixels to 5000X1000 pixels) and aspect ratios. I want to use neural networks for dealing with this problem. Is there any ...
  • 566
1 vote
1 answer
33 views

Image classification using Binary Cross Entropy but with only training examples for one of the classes e.g. class 1 VS anything else

I am training a 'specialist network' to reconstruct images of an object using a Variational Autoencoder (VAE). The training set (~15000 images) is of a single object in multiple poses. I also want ...
  • 23
1 vote
0 answers
107 views

YOLO v2 loss function

I'm trying to understand (and implement) the YOLOv2 loss function, which is not given explicitly in the original paper. There are several posts on this topic, but quite a few seem to confuse the ...
  • 11
0 votes
0 answers
14 views

Resizing images for object detection using Faster R-CNN

I would like to detect several largest objects and extract features from images using Faster R-CNN. Given that Faster R-CNN was trained on images with the shorter side of 600 pixels. Does it make ...
  • 2,511
0 votes
0 answers
23 views

Using Inception and FID scores in training?

Is it possible to use the Inception and FID scores in the training of a deep image generation model, i.e. to maximize the scores in a loss function, albeit this is "cheating"? If so, has ...
  • 1
1 vote
0 answers
69 views

Balancing Multiple Evaluation Metrics for a Model

When evaluating a machine learning (or other statistical model) against multiple evaluation metrics, is there a standardized way to choose the "best" model? As a concrete example, for a two ...
0 votes
0 answers
24 views

Best way to approximate head point having only face keypoints

I'm using the BlazeFace model from TensorFlow which only has this few keypoints: I need those keypoints plus a head keypoint, like this one: My question is, which would be the best way to ...
0 votes
0 answers
7 views

Structure from motion approach to estimate target frame from nearby frames

My question comes from the paper: https://arxiv.org/pdf/1704.07813.pdf , which is an unsupervised learning approach for depth estimation. Suppose I have a sequence of images $I_{t-1}, I_t, I_{t+1}$. I ...
  • 1
0 votes
0 answers
80 views

Is good to use binary images for deep learning?

I am trying to make a solar event detector from spectrograms. My problem is, I don't know how to approach the problem properly. I have generated high contrast images from the original data source in ...
  • 1
1 vote
0 answers
10 views

Finding a dataset for a computer vision project related to medical imaging (related to cancer/tumor) [closed]

I am trying to find a dataset of medical images related to tumor/cancer, there should be images different stages of the cancer and also preferably the details about the patient, their medical history, ...
0 votes
3 answers
174 views

How to calculate Precision and Recall for an image classification problem?

I'm not understanding how to calculate Precision and Recall if I'm doing image classification. If I have two classes, Cat and Dog, and for evaluation I get an image of a Dog and the model classifies ...
's user avatar
0 votes
0 answers
9 views

Image-recognition model makes good predictions only with training examples

Im trying to use a kaggle dataset to train a model that recognizes american fingerspelling language from an image. The problem is that, built the model, if i record the screen with the examples ...
0 votes
0 answers
14 views

Is classification with the backbone network a justifiable surrogate task in object detection?

There is a huge dataset of pictures but just a portion of the samples has bounding boxes. The main goal is to build an object detection model on the labeled dataset. One problem is that the dataset is ...
1 vote
0 answers
33 views

What is the difference between these two types of training?

Suppose that I want to detect if a picture contains a particular logo, for instance the following one. Since template matching would be slow and fail those scaled or resized ones, I decided to train ...
  • 5,446
0 votes
0 answers
12 views

Unconventional pretext task in computer vision - can I somehow justify it?

I was working on a industrial object detection neural network project. Since we had multiple images of the same object in different (but fixed) positions and light conditions, our dataset was very ...
0 votes
0 answers
13 views

Extract Local Attention Positions

I am working on a project, I want to extract the local attention positions present in the feature maps of every level in the FPN. Below is the visualization: I am using MMDetection toolbox for ...
0 votes
0 answers
49 views

Math behind predicting the dimensions of bounding boxes

I have been trying to understand the mechanics behind YOLO v3. I got stuck at the section which defines the calculation of bounding box. Going through various blogs, I found that The dimensions of ...
  • 101
4 votes
1 answer
2k views

Adam is an adaptive learning rate method, why people decrease its learning rate manually?

Adam optimizer is an adoptive learning rate optimizer that is very popular for deep learning, especially in computer vision. I have seen some papers that after specific epochs, for example, 50 epochs, ...
  • 217
0 votes
1 answer
48 views

Classification with Augmentation vs Contrastive Learning

How Contrastive Learning based on (SimCLRv2 approach) is compared to regular classification (VGG, Resnset, etc) with Data Augmentation. It seems to me, that it should have very similar performance. I'...
  • 615
0 votes
0 answers
36 views

How do you scale the activation function of an auto-encoder when using a custom normalization fitted on the data?

I'm working on a convolutional auto encoder. The input is an image The output is a reconstructed image During the training phase, we feed the same image in and out The loss is the Mean Squared Error ...
  • 111
1 vote
0 answers
312 views

Multi-Head Attention in ViT

I need help to understand the multihead attention in ViT. Here's the code I found from GitHub: ...
0 votes
0 answers
10 views

shape of output in feature pyramid network for RetinaNet

In FPN, each pyramid outputs a tensor which will go to classification and regression. This has an output of 256-d channels. But what is the complete shape of the output, including mini-batch size, ...
1 vote
1 answer
150 views

Is Intersection over Union in object detection differentiable?

Is IoU in object detection differentiable or can be back propagated? Is it used in the training process or just in the inference?
1 vote
0 answers
29 views

Annotations and Bounding Box Definition in Object Detection

I'm having confusion about the concepts of object detection. If the dataset has been annotated to have ground truth in (x, y) of the four vertices, then why are the papers (using the same dataset) are ...

1
2 3 4 5
10