Source-linked AI summary
DSSD : Deconvolutional Single Shot Detector
Cheng-Yang Fu, Wei Liu, Ananth Ranga, Ambrish Tyagi, Alexander C. Berg
TL;DR
The paper addresses how to introduce additional context into single-shot object detection, especially for small objects. It combines Residual-101 with SSD and adds deconvolution, feed-forward, and output modules to form DSSD. DSSD achieves state-of-the-art results on PASCAL VOC and COCO, while remaining comparable in speed to other detectors.
Problem
The paper seeks to improve general object detection by adding stronger features and large-scale context, particularly for small objects.
Method
DSSD combines SSD with Residual-101 and adds deconvolution layers, feed-forward connections, and a new output module.
Results
DSSD achieves state-of-the-art detection results on PASCAL VOC and COCO, including 33.2% mAP on COCO and 80.0% mAP on VOC2012 test.
Takeaways & Limitations
The encoder-decoder context approach improves detection, especially for small or context-specific objects, and can be applied beyond SSD to other detection methods.
Takeaways & Limitations
DSSD is slower than original SSD because of the deeper Residual-101 network, added modules, and more default boxes.
Abstract
from arXiv · showhide
The main contribution of this paper is an approach for introducing additional context into state-of-the-art general object detection. To achieve this we first combine a state-of-the-art classifier (Residual-101[14]) with a fast detection framework (SSD[18]). We then augment SSD+Residual-101 with deconvolution layers to introduce additional large-scale context in object detection and improve accuracy, especially for small objects, calling our resulting system DSSD for deconvolutional single shot detector. While these two contributions are easily described at a high-level, a naive implementation does not succeed. Instead we show that carefully adding additional stages of learned transformations, specifically a module for feed-forward connections in deconvolution and a new output module, enables this new approach and forms a potential way forward for further detection research. Results are shown on both PASCAL VOC and COCO detection. Our DSSD with $513 \times 513$ input achieves 81.5% mAP on VOC2007 test, 80.0% mAP on VOC2012 test, and 33.2% mAP on COCO, outperforming a state-of-the-art method R-FCN[3] on each dataset.
1. Introduction
The paper introduces DSSD to add large-scale context to object detection by combining Residual-101 with SSD and adding deconvolution layers. It motivates this design through improved feature networks, contextual information, and the particular needs of small-object detection, while noting that naive modifications require careful integration.
- DSSD combines the Residual-101 classifier with SSD and augments it with deconvolution layers to introduce additional context for detection.The approach targets accuracy improvements, especially for small objects.
- Recent object detection research has revisited sliding-window techniques as stronger deep-learning frameworks reduce the number of candidate boxes that must be considered.
- Better feature networks and additional context are identified as natural ways to improve detection accuracy, particularly for small objects.
- Figure 1 distinguishes SSD layers from the additional DSSD layers within the residual-network architectures.
- Residual-101 and deconvolution layers do not work out of the box, so the model requires carefully constructed combination and output modules.
2. Related Work
Related detection methods improve prediction by using multiple feature layers, scale-specific layers, or contextual regions. DSSD follows encoder-decoder hourglass designs to pass context into detection while addressing feature-resolution reduction.
- Many object detectors rely on a top ConvNet layer to detect objects at different scales, burdening one layer with modeling diverse sizes and shapes.
- Feature-combination methods make representations more descriptive for localization and classification but increase memory use and reduce speed.
- Figure 2 presents variants of the prediction module used in the model’s design.
- Scale-specific methods use layers with different receptive fields to predict objects at corresponding scales, including SSD’s multi-layer default-box predictions.
- DSSD uses an encoder-decoder hourglass structure so deconvolution layers both restore feature-map resolution and bring contextual information into prediction.
3. Deconvolutional Single Shot Detection (DSSD) model
DSSD builds on SSD by using Residual-101, prediction modules, and a shallow deconvolutional decoder with skip connections. The architecture adds contextual information while preserving a fast-detector design and uses SSD-style training and data-derived box ratios.
- SSD foundation: SSD predicts scores and box offsets for predefined default boxes using progressively smaller convolutional layers added to a base network.
- Residual-101 base network: DSSD replaces VGG with Residual-101, but the resulting SSD baseline reaches 76.4% mAP on PASCAL VOC 2007 test, below VGG SSD’s 77.5%.
- Prediction module: A residual block is added before each prediction layer to improve the effectiveness of prediction when using Residual-101.
- Deconvolutional SSD: Deconvolution layers progressively increase feature-map resolution, while skip connections integrate information from earlier feature maps.
- Deconvolutional SSD: The decoder is intentionally shallow because symmetric hourglass layers would double inference time and lack classification-pretrained decoder weights.
- Training: Training follows SSD by matching default boxes to ground-truth boxes, selecting negatives at a 3:1 ratio, and minimizing localization and confidence losses.
- Training: Seven aspect-ratio clusters are selected from training bounding boxes using K-means after increasing the cluster count when error improves by more than 20%.
4. Experiments
Experiments evaluate Residual-101-based SSD and DSSD on PASCAL VOC and COCO, including ablations of prediction and deconvolution modules. DSSD improves detection accuracy, particularly for small objects and contextual classes, while trading off speed against the original SSD.
- PASCAL VOC 2007: DSSD321 and DSSD513 are consistently about 1–1.5% better than corresponding SSD models without deconvolution layers on VOC2007.The authors attribute this comparison to adding deconvolution layers and skip connections.
- Ablation Study on VOC2007: Prediction modules improve Residual-101 SSD training, while stacking two modules produces little difference compared with one.The module is intended to prevent objective-function gradients from flowing directly into the Residual network backbone.
- Ablation Study on VOC2007: 78.6% mAP is obtained by the elementwise-product feature combination when adding the deconvolution module.Approximate bilinear pooling slowed training and reduced the rate of training-error decrease, so it was not evaluated further.
- PASCAL VOC 2012: 80.0% mAP is achieved on PASCAL VOC2012 without extra training data, multiple cropping, or an ensemble method during testing.The model is reported to outperform the other evaluated methods.
- COCO: DSSD513-Residual-101 achieves 33.2% mAP on COCO, 3.3% better than R-FCN, with improvements across all object sizes.The authors note that increasing input size may benefit DSSD further, but with longer training and inference time.
- Speed: DSSD is slower than original SSD because Residual-101 and the added prediction and deconvolution layers introduce computational overhead.Compared with R-FCN, DSSD513 has better accuracy but is slightly slower, while DSSD321 retains a speed advantage with a small accuracy drop.
5. Conclusion
DSSD adds context to SSD through an encoder-decoder hourglass design and demonstrates effectiveness on benchmark datasets. It outperforms SSD, particularly for small or context-specific objects, while preserving comparable speed to other detectors.
- DSSD adds context to a state-of-the-art object detection framework using an encoder-decoder hourglass approach.
- DSSD outperforms the previous SSD framework, especially on small objects or context-specific objects.
- DSSD achieves state-of-the-art detection results on PASCAL VOC and COCO.
- The approach preserves comparable speed to other detectors and could also be applied to R-CNN-series methods.