DSOD: Learning Deeply Supervised Object Detectors from Scratch

By Zhiqiang Shen, Zhuang Liu, Jianguo Li, Yu-Gang Jiang, Yurong Chen, Xiangyang Xue. 2017

628 citations recorded by the source corpus.

We present Deeply Supervised Object Detector (DSOD), a framework that can learn object detectors from scratch. State-of-the-art object objectors rely heavily on the off the-shelf networks pre-trained on large-scale classification datasets like Image Net, which incurs learning bias due to the difference on both the loss functions and the category distributions between classification and detection tasks. Model fine-tuning for the detection task could alleviate this bias to some extent but not fundamentally. Besides, transferring pre-trained models from classification to detection between discrepant domains is even more difficult (e.g. RGB to depth images). A better solution to tackle these two critical problems is to train object detectors from scratch, which motivates our proposed DSOD. Previous efforts in this direction mostly failed due to much more complicated loss functions and limited training data in object detection. In DSOD, we contribute a set of design principles for training object detectors from scratch. One of the key findings is that deep supervision, enabled by dense layer-wise connections, plays a critical role in learning a good detector. Combining with several other principles, we develop DSOD following the single-shot detection (SSD) framework. Experiments on PASCAL VOC 2007, 2012 and MS COCO datasets demonstrate that DSOD can achieve better results than the state-of-the-art solutions with much more compact models. For instance, DSOD outperforms SSD on all three benchmarks with real-time detection speed, while requires only 1/2 parameters to SSD and 1/10 parameters to Faster RCNN.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Radiology, Nuclear Medicine and Imaging

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Radiology, Nuclear Medicine and Imaging

Citations in the corpus

Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adaptive Training Sample Selection (2019, background): “Thereafter, plenty of works are presented to boost its performance in different aspects, such as fusing context information from different layers [24, 12, 69], training from scratch [50, 73], introducing new loss function [33, 6], anchor refinement and matching [66, 67], architecture redesign [21, 2”
DropBlock: A regularization method for convolutional networks (2018, background): “Recently, a few papers tried to address the issue using novel model architecture [33], large minibatch size [34], and better normalization layer [35].”
CornerNet: Detecting Objects as Paired Keypoints (2018, neutral): “4 DSOD300 (Shen et al., 2017a) DS/64-192-48-1 29.”
Rethinking the Value of Network Pruning (2018, methodology): “With a slight abuse of notation, here Scratch-E/B denotes "train the small model on classiﬁcation, transfer to detection", and is different from the setup of detection without ImageNet pre-training as in Shen et al. (2017).”
CenterNet: Keypoint Triplets for Object Detection (2019, neutral): “4 DSOD300 [35] DS/64-192-48-1 300×300 300×300 29.”
Bottom-Up Object Detection by Grouping Extreme and Center Points (2019, background): “One-stage object detector One-stage object detectors [22, 25, 28, 38, 39, 42, 48] do not have a region cropping module.”
DetNet: A Backbone network for Object Detection (2018, background): “DSOD [28] first proposes to train detection from scratch, whose results are lower than pretrained methods.”
Photographic Text-to-Image Synthesis with a Hierarchically-Nested Adversarial Network (2018, methodology): “Leveraging hierarchical representations of CNNs is an effective way to enhance implicit multi-scaling and ensembling for tasks such as image recognition [21, 48] and pixel or object classification [41, 25, 22, 37, 44, 50].”
Deep Feature Pyramid Reconfiguration for Object Detection (2018, neutral): “5.0 94.6 63.3 85.9 65.6 93.0 88.5 87.8 86.4 57.4 85.2 73.4 87.8 76.8 YOLOv2[38] Darknet-19 75.4 86.6 85.0 76.8 61.1 55.5 81.2 78.2 91.8 56.8 79.6 61.7 89.7 86.0 85.0 84.2 51.2 79.4 62.9 84.9 71.0 DSOD[42] DenseNet 76.3 89.4 85.3 72.9 62.7 49.5 83.6 80.6 92.1 60.8 77.9 65.6 88.9 85.5 86.8 84.6 51.1 7”
Deep Learning for Generic Object Detection: A Survey (2018, background): “number of recent approaches improve detection by predicting objects of different resolutions at different layers and then combining these predictions: SSD [175] and MSCNN [24], RBFNet [173], and DSOD [242]. SSD [175] spreads out default boxes of different scales to multiple layers within a CNN, and ”
Dynamic Anchor Feature Selection for Single-Shot Object Detection (2019, background): “For example, semantic information on different layers is enhanced in [9, 31, 37, 23] to boost the”
Object Detection With Deep Learning: A Review (2018, background): “The evaluated approaches include R-CNN [15], SPP-net [64], Fast R-CNN [16], networks on convolutional feature maps (NOC) [114], Bayes [85], MR-CNN& S-CNN [105], Faster R-CNN [17], HyperNet [101], ION [95], MS-GR [104], StuffNet [100], SSD300 [71], SSD512 [71], OHEM [113], SDP + CRC [34], G-CNN [70],”
Understanding and Controlling User Linkability in Decentralized Learning (2018, methodology): “All images in both datasets are publicly available on Flickr and have been used in various prior works [44, 51, 63, 72, 83].”
Gradient-Leaks: Understanding and Controlling Deanonymization in Federated Learning (2018, methodology): “his makes them ideal to train a classifier on examples resembling usergenerated data on clients. All images in both datasets are publicly available on Flickr and have been used in various prior works [44, 51, 63, 72, 83]. 7.1 PIPA PIPA [82] is a large-scale dataset of ∼37k personal photos uploaded b”
A Survey of Deep Learning-Based Object Detection (2019, background): “Some works [115]–[118] train object detectors from scratch by dense layer-wise connections.”