Source-linked AI summary
G-CNN: an Iterative Grid Based Object Detector
Mahyar Najibi, Mohammad Rastegari, Larry S. Davis
TL;DR
Object detection methods often rely on computationally costly proposal stages, creating a need for accurate detection without proposals. G-CNN iteratively moves and scales a fixed multi-scale grid toward objects using piecewise regression, achieving comparable Fast R-CNN performance with fewer boxes and faster detection.
Problem
Object proposal generation is a major computational bottleneck in CNN-based object detection, while removing it has previously caused performance degradation.
Method
G-CNN starts from a fixed multi-scale grid and uses piecewise regression over multiple steps to move boxes toward objects and classify them.
Results
G-CNN achieves comparable state-of-the-art detection results to Fast R-CNN on PASCAL VOC, is about 5X faster, and can surpass Fast R-CNN with around 180 boxes versus around 2K proposals.
Takeaways & Limitations
Removing the proposal stage can reduce the boxes processed while preserving comparable detection performance.
Takeaways & Limitations
Training uses an approximate update that assumes the previous-step regressor could be learned perfectly, making the loss computation tractable.
Abstract
from arXiv · showhide
We introduce G-CNN, an object detection technique based on CNNs which works without proposal algorithms. G-CNN starts with a multi-scale grid of fixed bounding boxes. We train a regressor to move and scale elements of the grid towards objects iteratively. G-CNN models the problem of object detection as finding a path from a fixed grid to boxes tightly surrounding the objects. G-CNN with around 180 boxes in a multi-scale grid performs comparably to Fast R-CNN which uses around 2K bounding boxes generated with a proposal technique. This strategy makes detection faster by removing the object proposal stage as well as reducing the number of boxes to be processed.
1. Introduction
G-CNN removes the computationally costly object-proposal stage by iteratively moving a multi-scale grid of boxes toward and classifying objects. It achieves comparable detection performance while using fewer boxes and improving speed.
- Object proposals are a major computational bottleneck in state-of-the-art detectors, while prior attempts to remove them degraded performance.
- G-CNN starts from a regularly sampled multi-scale grid and iteratively updates boxes toward objects while classifying their categories.Its piecewise regressor handles the nonlinear mapping through several easier regression steps.
- At test time, each box receives category confidence scores and is updated using the regressor for its most probable class.
- G-CNN achieves state-of-the-art Fast R-CNN results on PASCAL VOC without computing bounding-box proposals and is about 5X faster.
2. Related Work
CNN-based detectors improved object detection but commonly rely on computationally expensive proposal stages and process many candidate boxes. Related approaches reduce or replace proposals, yet introduce efficiency or accuracy trade-offs that motivate G-CNN.
- R-CNN uses object proposals, applies a CNN to each proposed box, and classifies the object category inside it.
- Around 2K proposals create redundant feature computation in R-CNN; Fast R-CNN shares convolutional features but still excludes proposal generation from its unified network.
- Multibox outputs 800 boxes but expands the CNN’s final layer, while Faster-RCNN still starts from thousands of anchor points to propose 300 boxes.
- AttentionNet applies iterative regression for single-category detection, but its multiple-object use generates thousands of proposals and reapplies the model inefficiently.
- Fixed-box approaches challenge the necessity of proposals, but comparable performance can require a distribution derived from proposal methods.
3. G-CNN Object Detector
G-CNN treats detection as iterative search from a fixed multi-scale grid toward assigned object boxes. Its piecewise regressor updates boxes over multiple steps, while shared global features reduce test-time computation.
- Iterative detection: G-CNN trains a CNN to move and scale fixed multi-scale grid boxes toward objects through iterative regression.The method begins with a spatial pyramid of boxes and learns successive changes toward assigned targets.
- Iterative detection: The iterative approach makes nonlinear long-range localization tractable by learning small box changes step by step.The target box advances along the path to the assigned ground-truth box until the final training step.
- Training targets: Training assigns each initial grid box with IoU above 0.2 to a ground-truth box, and that assignment remains fixed throughout training.The assignment is many-to-one and uses the ground-truth boxes in the same image.
- Regression model: The regressor predicts four class-specific values encoding scale and translation changes for each box.Width and height use log-scale shifts, while translation is scale invariant; smooth l1 loss is used for regression.
- Test-time execution: At test time, convolutional global features are computed once, while the regression part repeatedly produces box modifications.This decomposition avoids recomputing all network layers inside the iterative detection loop.
- Test-time execution: With Stest = 5 and approximately 180 grid boxes, G-CNN achieves results comparable to Fast R-CNN using approximately 2K proposals.G-CNN performs the regression forward calculation Stest × Ngrid times instead of processing Nproposal proposals.
4.1. Experimental Setup
The experiments evaluate G-CNN on Pascal VOC 2007 and 2012 using AlexNet and VGG16 backbones under a standardized image-scaling setup.
- Datasets and models: Experiments use the Pascal VOC 2007 and Pascal VOC 2012 object-detection datasets.Performance is evaluated with both a small AlexNet CNN and a very deep VGG16 CNN.
- Datasets and models: G-CNN is evaluated with AlexNet as a small CNN structure and VGG16 as a very deep CNN structure.Both models are pretrained with weights learned from ImageNet.
- Image preprocessing: Images are scaled so the shortest side is 600 pixels while the longer side does not exceed 1000 pixels, preserving aspect ratio.The stated scaling rule can result in a shortest side below 600 pixels when the aspect ratio constraint applies.
- Grid configuration: Training uses an initial overlapping spatial pyramid with [2] scales and overlaps of [0.9,0.8,0.7] across scales.Overlap α defines horizontal and vertical strides as widthcell ∗ (1 − α) and heightcell ∗ (1 − α).
4.2. Results on VOC datasets
On VOC datasets, G-CNN replaces proposal algorithms with a fixed grid while achieving results comparable to Fast R-CNN under the reported evaluation settings.
- VOC evaluation: G-CNN replaces object proposals with a fixed multi-scale grid for evaluating detection performance.The comparison fixes the classifier to Fast R-CNN and uses selective search for the original Fast R-CNN baseline.
- VOC2007: Around 180 initial boxes are enough for G-CNN to surpass Fast R-CNN, which uses around 2K selective-search boxes, on VOC2007.This is the G-CNN(5) configuration with five test-time regression steps and reduced grid overlaps.
- VOC2012: On VOC2012, G-CNN obtains almost the same result as Fast R-CNN when both methods are trained on VOC2012 trainval.The reported Fast R-CNN comparison uses a slightly different training-data condition noted in the passage.
4.3. Stepwise training matters
The experiments test whether G-CNN’s stepwise training is necessary by comparing it with iterative Fast R-CNN regression and one-step grid training.
- Training strategies: The ablation compares G-CNN with IF-RCNN and 1Step-Grid to assess whether stepwise training can be replaced by simpler training.1Step-Grid uses the same training tuples as G-CNN but trains the regressor in one step.
- Experimental setup: All methods use AlexNet, the same initial grid, five test-time iterations, and VOC2007 trainval/test splits.Figure 4 compares G-CNN and IF-RCNN across different numbers of regression steps.
- Findings: G-CNN outperforms 1Step-Grid by a considerable margin despite using the same training samples.The result identifies stepwise training as crucial to G-CNN’s success.
4.4. Analysis of the detection results
The analysis examines whether removing proposals harms localization and compares false-positive patterns across G-CNN, Fast R-CNN, and 1Step-Grid.
- Localization analysis: Removing the proposal stage did not hurt G-CNN’s localization relative to Fast R-CNN.For the furniture class, G-CNN slightly improved false positives caused by localization error.
- Training effects: 1Step-Grid has a higher rate of localization false positives than G-CNN despite using the same training tuples.This provides another indication that G-CNN’s multi-step training strategy matters.
- False positives: Figure 5 categorizes top-ranked false positives as localization errors, similar-class confusion, background errors, or other sources.Localization errors include duplicate detections or detections with IoU between 0.1 and 0.5.
4.5. Qualitative results
Qualitative examples visualize G-CNN’s paths through bounding-box space as boxes change position and scale from an initial three-scale grid.
- Bounding-box paths: G-CNN changes box position and scale to fit different objects, producing both successful and failure paths.Figure 6 shows successful examples in its first four rows and failure examples in the last rows.
4.6. Detection run time
G-CNN avoids proposal generation and processes a much smaller grid, yielding substantially lower reported detection time than Fast R-CNN under the described setup.
- G-CNN's timing comparison uses truncated SVD compression for fc6 and fc7 layers in both methods.The retained singular-value counts are 1024 for fc6 and 256 for fc7.
- Both methods use the VGG16 network, and G-CNN uses the same classifier as Fast R-CNN.Timings were performed on a system with two K40 GPUs.
- 2050 ms/im is the total Fast R-CNN detection time, including 1830 ms for Selective Search and 220 ms for object detection.Selective Search generates around 2K bounding boxes and is not implemented in GPU mode.
- 363 ms/im is the reported total G-CNN detection time using five iterations over around 180 boxes.The global network takes 188 ms per image, while each segmentation iteration takes 35 ms; classification runs in parallel.
5. Conclusion
G-CNN formulates object detection as iterative search from fixed boxes toward image objects, removing proposal generation while achieving comparable detection results and faster runtime.
- G-CNN models detection as an iterative search over possible bounding boxes, migrating a fixed grid toward objects in the image.The method uses piecewise regression to move boxes step by step through the nonlinear search problem.
- Removing the object proposal stage is the method's main contribution because proposal generation is a bottleneck in CNN-based detection systems.The paper identifies proposal techniques as a major computational bottleneck for state-of-the-art detectors.
- 5X faster than Fast R-CNN, G-CNN achieves comparable results to state-of-the-art detectors.The conclusion reports both the runtime advantage and comparable detection performance.
- G-CNN can fail through classifier firing on similar objects and poor localization when similar objects overlap heavily.Figure 6 illustrates successful and failed paths on the VOC2007 test set using AlexNet.