Segmentation and Recognition Using Structure from Motion Point Clouds
read more
Citations
SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation
The SYNTHIA Dataset: A Large Collection of Synthetic Images for Semantic Segmentation of Urban Scenes
nuScenes: A multimodal dataset for autonomous driving
ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation
BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation
References
Rapid object detection using a boosted cascade of simple features
Multiple view geometry in computer vision
Multiple View Geometry in Computer Vision.
A Combined Corner and Edge Detector
Extremely randomized trees
Related Papers (5)
Frequently Asked Questions (10)
Q2. What have the authors stated for future works in "Segmentation and recognition using structure from motion point clouds" ?
The authors hope that semi-supervised techniques that use extra partially labeled or unlabeled training data may lead to improved performance in the future.
Q3. How did the authors train the randomized decision forest?
The authors trained a randomized decision forest based on their five motion and structure cues, using combined day and dusk sequences for both training and testing.
Q4. What are the 11 categories of the labeled data?
The labeled data has 11 categories: Building, Tree, Sky, Car, Sign-Symbol, Road, Pedestrian, Fence, Column-Pole, Sidewalk, and Bicyclist.
Q5. How can the algorithm be trained on point clouds?
By including a fixed offset Cy, the algorithm can be trained on point clouds from one vehicle, but run on other cameras and vehicles.
Q6. What is the effect of balancing the categories?
One by-product of balancing the categories during training is that the areas of smaller classes in the images tend to be overestimated, spilling out into the background (e.g., the bicycle in Fig. 7).
Q7. What is the way to model the joint dependencies of the two classifiers?
Learning a histogram for each pair of (motion and structure, appearance) tree leaf nodes could better modelthe joint dependencies of the two classifiers, but care must be taken so that in avoiding overfitting, quadratically more training data is not required.
Q8. What are the advantages of motion and structure features over appearance features?
11Motion and structure features do however have an obvious advantage over appearance features: generalization to novel lighting and weather conditions.
Q9. How did the authors determine the relative importance of the motion and structure cues?
To determine the relative importance of the five motion and structure cues, the authors analyzed the proportion of each chosen by the learning algorithm, as a function of depth in the randomized forest.
Q10. What are the common types of images that are used to label?
Existing databases of labeled images do not include frames taken from video sequences, and usually label relevant classes with only bounding boxes.