Source-linked AI summary
DenseNet: Implementing Efficient ConvNet Descriptor Pyramids
Forrest Iandola, Matt Moskewicz, Sergey Karayev, Ross Girshick, Trevor Darrell, Kurt Keutzer
TL;DR
Dense CNN detection can be prohibitively slow because many overlapping regions repeat convolutional work. DenseNet computes reusable, dense multiscale CNN feature pyramids, achieving 1sec extraction for a 25-scale pyramid versus 10sec for 2000 traditional window proposals, while qualitatively approximating isolated region descriptors.
Problem
Dense CNN detection can be prohibitively slow because many overlapping regions repeat convolutional work.
Method
DenseNet computes dense, multiscale CNN features that support reuse across overlapping region proposals.
Results
1sec for a 25-scale feature pyramid compared with 10sec for 2000 traditional AlexNet window proposals on an NVIDIA K20 GPU.
Takeaways & Limitations
DenseNet’s dense descriptors appear visually similar to descriptors computed independently on cropped regions.
Takeaways & Limitations
Stitched-image processing can introduce edge artifacts and receptive-field pollution between neighboring pyramid scales, requiring padding to mitigate these effects.
Abstract
from arXiv · showhide
Convolutional Neural Networks (CNNs) can provide accurate object classification. They can be extended to perform object detection by iterating over dense or selected proposed object regions. However, the runtime of such detectors scales as the total number and/or area of regions to examine per image, and training such detectors may be prohibitively slow. However, for some CNN classifier topologies, it is possible to share significant work among overlapping regions to be classified. This paper presents DenseNet, an open source system that computes dense, multiscale features from the convolutional layers of a CNN based object classifier. Future work will involve training efficient object detectors with DenseNet feature descriptors.
1 Introduction
CNNs improve object classification, but extending them to dense sliding-window detection can make computation prohibitively slow. DenseNet addresses this bottleneck by sharing computation across overlapping regions through dense, multiscale CNN features.
- Object detection combines object localization with classification, unlike classification systems that receive object locations or crops.
- Sliding-window detectors simplify region proposal generation, but dense position-and-scale sampling can produce very large proposal sets.
- Around 33% mAP was typical for strong sliding-window detectors on PASCAL 2007, with hand-engineered descriptors providing incremental improvements.
- At approximately 50ms per region and 200K regions, naive dense CNN detection would require approximately 3 hours per image.
- DenseNet computes dense, multiscale CNN descriptor pyramids and targets efficient classification across many possible region proposals.
2 Related Work
Prior detectors either use dense CNN pyramids or region proposals, but both expose accuracy, localization, runtime, or generality challenges. DenseNet targets reuse of shared convolutional computation across overlapping proposals.
- DetectorNet uses a coarse three-scale CNN pyramid and reported 30% mAP on PASCAL VOC 2007 without ImageNet pretraining.
- OverFeat generates dense, multiscale CNN features, while the cited discussion contrasts its platform focus with DenseNet’s intended generality and GPU efficiency.
- R-CNN combines class-independent region proposals with CNN descriptors and achieved 54% mAP on PASCAL 2007, increasing to 59% with bounding-box regression.
- R-CNN’s approximately 10-second image latency reflects repeated processing of overlapping proposals, motivating shared convolutional computation.
- DenseNet reuse is complicated by per-region data centering and proposals with differing sizes and aspect ratios.
3 DenseNet CNN Feature Pyramids
DenseNet constructs multiscale CNN feature pyramids while addressing batch-processing constraints, data centering, receptive-field pollution, and varied region geometry. It provides open-source interfaces and substantially faster pyramid extraction than independent window processing.
- DenseNet exploits translational invariance after centering: overlapping same-size, same-aspect-ratio regions can share convolutional descriptors.
- The method replaces mean-image centering with a single mean pixel because per-region mean-image centering is unclear for stitched inputs.
- The method builds configurable image pyramids to handle differing region scales and delegates most aspect-ratio handling to later detector stages or separate warped pyramids.
- Pyramids typically contain approximately 10–50 levels and 3–8 times the original image’s total pixels.
- DenseNet stitches pyramid scales into large images for efficient Caffe processing, mitigating receptive-field contamination with 16-pixel borders and interpolated padding.
- 0.2% lower top-1 accuracy resulted from one RGB mean pixel versus an RGB mean mask on pretrained AlexNet ImageNet classification.
- 3.4 Measured Speedup: 1sec for a 25-scale feature pyramid compared with 10sec for 2000 traditional AlexNet window proposals on an NVIDIA K20 GPU.
- DenseNet supplies Matlab and Python pyramid-extraction APIs integrated with the open-source Caffe framework and designed to interoperate with DPM HOG code.
4 Qualitative Evaluation
The qualitative evaluation compares descriptors extracted independently from cropped regions with descriptors cropped from a dense feature pyramid. Their visual similarity supports DenseNet as an approximation to isolated per-region descriptors.
- The baseline crops regions in pixel space and computes each CNN descriptor independently, which is computationally inefficient for many regions.
- DenseNet first computes descriptors densely without proposal windows, then crops corresponding regions from its descriptor pyramids.
- Descriptors from the independent and DenseNet pipelines look similar in the visualized examples, suggesting DenseNet approximates isolated per-region descriptors.
5 Conclusion
DenseNet computes descriptors densely on a full image, then crops descriptor regions to approximate region proposals rather than computing each region independently.
- Independent region descriptor computation first crops image regions and then computes descriptors separately for each region.The example uses arbitrarily chosen square regions, so pixel warping is unnecessary.
- DenseNet computes descriptors on full images before cropping descriptor regions to approximate region proposal windows.This extraction strategy is the type of feature computation DenseNet is optimized for.
- The two pipelines differ in whether descriptor computation precedes or follows region selection.DenseNet computes descriptors densely without regard to proposal windows, whereas the baseline processes each window independently.