Research map / Detection and segmentation / CNN features
Exploring Salient Object Detection with Adder Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adapts adder neural networks (AdderNet-50) to salient object detection, using a differential merging module and differential channel attention to fix feature-diversity loss in adder decoders, cutting energy about 55% versus CNN counterparts.
Problem
Adder neural networks are far more energy efficient than CNNs, but naively replacing convolutions with adder layers in salient object detection decoders causes large performance drops due to low feature diversity and unstable batch-norm statistics.
Approach
A differential merging module splits an adder layer into two branches, flips the sign of one, and takes their difference to shift output statistics near zero; a differential channel attention (DCA) built on the same idea replaces SE to rescale channels in adder networks.
Key results
- DMNet (AdderNet-50 + DMM + DCA) reaches maxF 0.751 / MAE 0.057 on DUT-O, versus 0.647 / 0.084 for the plain adder baseline.
- Applying DMM and DCA to PoolNet, FPN and PicaNet yields adder versions with comparable accuracy while saving around 70% energy.
- Adder versions of ICON and EDN achieve comparable performance to their CNN counterparts with 65% and 59% of energy preserved.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DMM instead of the plain ANN decoder (ablation) | AdderNet-50 SOD baseline · DUT-O | maxF: 0.647 -> 0.725 | Helped |
| Use DMM instead of the plain ANN decoder (ablation) | AdderNet-50 SOD baseline · PASCAL-S | maxF: 0.764 -> 0.813 | Helped |
| Add DCA to DMM (ablation) | DMNet · DUT-O | maxF: 0.725 -> 0.751 | Helped |
| Add DCA to DMM (ablation) | DMNet · PASCAL-S | maxF: 0.813 -> 0.833 | Helped |
| Increase decoder channel count (ablation) | ANN baseline · DUT-O | maxF: 0.647 -> 0.687 | Helped |
| Replace DCA with SE attention (ablation) | DMM · DUT-O | maxF: 0.751 -> 0.735 | Hurt |
| Use adder-based SE attention instead of DCA (ablation) | DMM · DUT-O | maxF: 0.751 -> 0.731 | Hurt |
| Set batch size to 128 rather than 64 (ablation) | Adder PoolNet · DUTS-TE | maxF: 0.855 -> 0.845 | Hurt |
| Increase initial learning rate from 2e-4 to 4e-4 (ablation) | Adder PoolNet · DUTS-TE | maxF: 0.855 -> 0.836 | Hurt |
| Train for more epochs (ablation) | Adder PoolNet · DUTS-TE | maxF: 0.825 -> 0.855 | Helped |
| Replace convolutional decoder operators with adder operators (ablation) | PoolNet · ECSSD | maxF: 0.933 -> 0.881 | Hurt |
| Replace convolutional decoder operators with adder operators (ablation) | PoolNet · DUT-O | maxF: 0.772 -> 0.733 | Hurt |
Surprising findings
- Directly replacing decoder convolutions with adder layers causes a large drop (e.g., PoolNet F-measure 0.833 to 0.747 on PASCAL-S), while replacing the encoder causes only a slight drop.
- Simply doubling decoder channels improves results less than the differential merging module, which adds fewer parameters and energy.
Limitations
- Performance still lags behind state-of-the-art CNN-based SOD models.
- Adder layers converge more slowly, requiring 120 epochs of training.
- Study limited to the AdderNet-50 backbone and FPN-style decoders; more complex architectures left for future work.
Experimental setup
Datasets: DUTS-TR
Benchmarks: ECSSD DUT-O PASCAL-S HKU-IS DUTS-TE
Metrics: max F-measure MAE S-measure energy (mJ)
Baselines: PoolNet FPN PicaNet CPD BASNet AFNet LDF GateNet-R ITSD MINet DCN CTDNet
Builds on: AdderNet PoolNet Feature Pyramid Network SE attention ICON EDN
Closest papers by meaning
- An Empirical Study of Adder Neural Networks for Object Detection (NeurIPS 2021 · similarity 0.91)
- Cascaded Partial Decoder for Fast and Accurate Salient Object Detection (CVPR 2019 · similarity 0.89)
- Salient Object Detection With Pyramid Attention and Salient Edges (CVPR 2019 · similarity 0.89)
- Deep Embedding Features for Salient Object Detection (AAAI 2019 · similarity 0.89)
- F³Net: Fusion, Feedback and Focus for Salient Object Detection (AAAI 2020 · similarity 0.88)
- Recurrently Aggregating Deep Features for Salient Object Detection (AAAI 2018 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.