The Pascal Visual Object Classes Challenge: A Retrospective
read more
Citations
Very Deep Convolutional Networks for Large-Scale Image Recognition
Very Deep Convolutional Networks for Large-Scale Image Recognition
ImageNet Large Scale Visual Recognition Challenge
Fully convolutional networks for semantic segmentation
You Only Look Once: Unified, Real-Time Object Detection
References
ImageNet Classification with Deep Convolutional Neural Networks
Distinctive Image Features from Scale-Invariant Keypoints
LIBSVM: A library for support vector machines
Histograms of oriented gradients for human detection
Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation
Related Papers (5)
Frequently Asked Questions (9)
Q2. What were the dominant methods used in the 2009 segmentation challenge?
The two dominant methods were: hierarchical random fields with a range of potentials, and the use of multiple bottom-up segmentations, combined with a classifier to predict the degree of overlap of a segment with an object.
Q3. What were the key extensions of the segmentation challenge?
Key extensions were:– use of multiple bottom-up segmentations to avoid making early incorrect boundary decisions, – Hierarchical MRFs e.g. modelling object cooccurrence, – use of parts-based instance models to refine detections, – deeper integration of segmentation model with detection/classification models, – use of 3D information.
Q4. What was the responsibility of the participant’s system to filter multiple detections of the same object?
Multiple detections of the same object in an image were considered false detections, e.g. 5 detections of a single object counted as 1 correct detection and 4 false detections – it was the responsibility of the participant’s system to filter multiple detections from its output.
Q5. What is the method the authors investigate for the super-classifier?
The method the authors investigate for the super-classifier is a linear classifier for each of the VOC classes, where the feature vector consists of the real-valued scores supplied by each submitted method.
Q6. What was the purpose of interpolating the precision-recall curve?
The intention in interpolating the precision-recall curve was to reduce the impact of the ‘wiggles’ in the precisionrecall curve, caused by small variations in the ranking of examples.
Q7. How can a normalised average precision measure be calculated?
A normalised average precision measure for detection can be computed by averaging normalised precisions computed at a range of recalls.
Q8. What is the way to avoid sliding windows search?
Another way of avoiding sliding windows search is to hypothesise bounding boxes bottom up, e.g. based on multiple segmentations (Van de Sande et al, 2011).
Q9. What is the way to measure the accuracy of a truncated object?
In general (data not shown) performance with respect to aspect ratio is better for lessextreme aspect ratios, and it is better for non-truncated objects than truncated ones (except that the top three methods in 2009 and 2012 all prefer truncated over nontruncated cats).