PaperlayerView exact revision on arXiv

Source-linked AI summary

From Image-level to Pixel-level Labeling with Convolutional Networks

Pedro O. Pinheiro, Ronan Collobert

arXiv:1411.6228v3cs.CV

TL;DR

Object segmentation requires pixel labels, but segmentation annotations are less abundant than classification labels. This paper trains a CNN using image-level labels to infer pixel-level classes, and reports large gains over prior weakly supervised methods alongside competitive results for several classes against fully supervised systems.

  • Problem

    Object segmentation requires pixel-level labels, while fully labeled segmentation data are less abundant than image-level classification labels.

    Source: §1 · pp. 1–2

  • Method

    A CNN is trained from image-level class labels with an aggregation constraint emphasizing classification-important pixels, then predicts labels densely at test time with few smoothing priors.

    Source: §1 · p. 2, §3.2.1 · p. 5

  • Results

    Average per-class accuracy increases from 30% to 90% over previous weakly supervised algorithms, and the model is competitive with fully supervised systems for several classes.

    Source: §4.3 · p. 7, §5 · p. 7

  • Takeaways & Limitations

    The results suggest that image-level annotations may help circumvent reliance on costly segmentation datasets for weakly supervised object segmentation.

    Source: §5 · p. 7

    Class performance varies with how closely ImageNet and Pascal VOC image statistics align, and Pascal VOC is used to select class-prior thresholds.

    Source: §4.3 · p. 7

Abstract

from arXiv · show

We are interested in inferring object segmentation by leveraging only object class information, and by considering only minimal priors on the object segmentation task. This problem could be viewed as a kind of weakly supervised segmentation task, and naturally fits the Multiple Instance Learning (MIL) framework: every training image is known to have (or not) at least one pixel corresponding to the image class label, and the segmentation task can be rewritten as inferring the pixels belonging to the class of the object (given one image, and its object class). We propose a Convolutional Neural Network-based model, which is constrained during training to put more weight on pixels which are important for classifying the image. We show that at test time, the model has learned to discriminate the right pixels well enough, such that it performs very well on an existing segmentation benchmark, by adding only few smoothing priors. Our system is trained using a subset of the Imagenet dataset and the segmentation experiments are performed on the challenging Pascal VOC dataset (with no fine-tuning of the model on Pascal VOC). Our model beats the state of the art results in weakly supervised object segmentation task by a large margin. We also compare the performance of our model with state of the art fully-supervised segmentation approaches.

1. Introduction

Object segmentation must assign class labels to pixels despite wide variation in object appearance and limited segmentation training data. The paper uses image-level labels to train a CNN to infer pixel labels, bridging classification and segmentation without segmentation annotations.

1. Introduction

  • The task is challenging because objects vary in position, pose, lighting, texture, shape, and background, while segmentation training data are limited.Source: §1 · p. 1
  • Image-level class labels are more abundant than segmentation labels, motivating a bridge that transfers learned features from classification to segmentation.CNNs can learn sufficiently general features for transfer across vision tasks.Source: §1 · pp. 1–2
  • The model uses only an image-level class label and treats segmentation as inferring the unknown pixels matching that label, a Multiple Instance Learning formulation.No segmentation labels or bounding-box annotations are used for training.Source: §1 · p. 2
  • An added aggregation layer constrains training to emphasize classification-important pixels; at test time it is removed and the CNN infers each pixel label.The network is trained by maximizing classification likelihood on an ImageNet subset.Source: Figure 1 · p. 1, §1 · p. 2

2. Related Work

Prior weakly supervised segmentation methods use structured superpixel models, while transfer-learning work applies ImageNet-trained CNNs to other vision tasks. This paper instead learns features for pixel-level segmentation from image-level annotations without fine-tuning on Pascal VOC.

2. Related Work

  • Earlier weakly supervised methods model superpixel relationships with a Multi-Image Model or learn spatial structure from graphlets and a Gaussian mixture prior.The Multi-Image Model connects superpixels across training images by appearance similarity; the graphlet approach transfers labels through manifold embedding.Source: §2 · p. 2
  • The proposed approach learns features for segmentation rather than designing task-specific features, and trains on ImageNet rather than the Pascal VOC evaluation data.The model is not fine-tuned on Pascal VOC or trained with segmentation or bounding-box information.Source: §2 · pp. 2–3
  • Other work adapts ImageNet-trained CNNs to Pascal VOC classification or detection, while SDS combines detection candidates with category-specific figure-ground predictions.The cited approaches include [10],, and [11].Source: §2 · pp. 2–3
  • Unlike prior transfer-learning work focused on classification or detection, this approach infers pixel labels using only image-level annotations.The approach uses ImageNet-learned features without refining the model on segmentation data.Source: §2 · pp. 2–3

3. Architecture

The architecture trains pixel-level class scores from image labels by aggregating them into image-level predictions, then uses image-level and smoothing priors at inference. The priors suppress false positives and promote locally consistent labels.

3. Architecture

  • A CNN built on frozen, ImageNet-trained Overfeat layers adds four convolutional layers to produce class and background score planes from image-level labels.For a 400 × 400 input, the architecture outputs scores on a 12-times-downsampled image grid.Source: §3 · p. 3

3.1. Multiple Instance Learning

  • The model aggregates pixel scores into image-level class scores, applies a softmax, and maximizes classification likelihood using image labels.Training backpropagates through the aggregation procedure and softmax.Source: §3.1 · p. 3
  • Log-Sum-Exp approximates max pooling smoothly, avoiding sum pooling’s equal weighting of all pixels and max pooling’s slow convergence.Its parameter controls the approximation’s smoothness and how similarly scored pixels contribute during training.Source: §3.1 · p. 4

3.2. Inference

  • At inference, the aggregation layer is removed, pixel scores become conditional probabilities, and shifted input images can recover labels across the full image.The image-level prior reweights pixel probabilities; local smoothing priors then promote consistent regions.Source: §3.2 · p. 4, §3.2.1 · p. 5

3.2.1 Adding Segmentation Priors

  • Smoothing priors suppress low-object-probability pixels as background and enforce local label consistency, reducing false positives while increasing true positives.The paper reports that more complex smoothing priors improve accuracy.Source: §3.2.1 · p. 5

4. Experiments

The experiments evaluate a segmentation model trained from ImageNet image labels on Pascal VOC, which was used for testing and threshold selection rather than feature training. The model substantially outperforms prior weakly supervised methods and is also compared with fully supervised systems.

4. Experiments

  • The experiments compare a model trained without segmentation labels against both weakly supervised and fully supervised segmentation systems.Pascal VOC 2012 validation data was used to choose architecture and hyperparameters, and to select class-confidence thresholds for smoothing priors.Source: §4 · p. 5, §4.2 · p. 6

4.1. Datasets

  • The model is trained on an ImageNet classification set of about 700,000 class images and 60,000 background images, while Pascal VOC is reserved for testing.Training images cover subclasses beneath the 20 target classes; the authors report difficulty finding images containing none of those classes.Source: §4.1 · p. 6, §4.2 · p. 6

4.2. Experimental Setup

  • Training uses a frozen Overfeat feature extractor followed by four convolutional layers, with 400 × 400 image patches as samples.The network has about 20 million parameters and uses dropout at rate 0.5 on each layer.Source: §4.2 · p. 6

4.3. Experimental Results

  • Average per-class accuracy rises from 30% to 90% against prior weakly supervised algorithms on Pascal VOC 2008, 2009, and 2010.The comparison uses the same metric and datasets as Zhang et al.; Table 2 also compares three aggregation layers.Source: Table 2 · p. 6, §4.3 · p. 7
  • Against fully supervised Pascal VOC 2012 systems, performance increases with more complex smoothing priors, while results are near state of the art for several classes and worse for others.Class statistics for some categories are closer between ImageNet and Pascal VOC than for others; selecting training images more similar to the test data could improve challenge results.Source: §4.3 · p. 7
  • Smoothing priors reduce false positives and smooth predictions; the Table 2 comparison across aggregation layers empirically supports the Log-Sum-Exp choice.Table 4 compares no prior, an image-level prior, and combinations with smoothing priors on Pascal VOC 2012 validation data.Source: §4.3 · p. 7

5. Conclusion

The framework segments objects with weak supervision and few simple priors, outperforming prior weakly supervised systems by a large margin. It is also competitive with fully supervised systems for several classes.

5. Conclusion

  • The model surpasses prior weakly supervised segmentation systems by a large margin and achieves competitive performance against fully supervised systems for at least several classes.Source: §5 · p. 7
  • The framework segments objects at the pixel level with weak supervision and only a few simple segmentation priors.Source: §5 · p. 7
  • Figure 4 contrasts outputs using the image-level prior with outputs using that prior plus SP-seg smoothing; Table 4 reports per-class average precision across prior settings on Pascal VOC 2012 validation.Source: Figure 4 · p. 8, Table 4 · p. 8

Related papers

This paper cites

Loading 1411.6228v3…