Early-Exit Criteria for Edge Semantic Segmentation
Saved in:
| Title: | Early-Exit Criteria for Edge Semantic Segmentation |
|---|---|
| Authors: | Gilbert, Mateus, Pacheco, Roberto, Couto, Rodrigo, Fladenmuller, Anne, Dias de Amorim, Marcelo, de Campos, Marcello, Campista, Miguel Elias |
| Contributors: | da Silva Gilbert, Mateus |
| Source: | 2025 IEEE International Conference on Machine Learning for Communication and Networking (ICMLCN). :1-6 |
| Publisher Information: | IEEE, 2025. |
| Publication Year: | 2025 |
| Subject Terms: | [INFO.INFO-CV] Computer Science [cs]/Computer Vision and Pattern Recognition [cs.CV], [INFO.INFO-NI] Computer Science [cs]/Networking and Internet Architecture [cs.NI], EE-DNN edge/cloud semantic segmentation, edge/cloud, EE-DNN, [INFO] Computer Science [cs], semantic segmentation |
| Description: | Early-exit deep neural networks (EE-DNNs) are multi-output DNNs designed for resource-constrained and latency-sensitive implementations, leveraging auxiliary output layers to partition processing between local, edge, and cloud devices. In the case of semantic segmentation, the literature lacks an efficient exit policy to stop the inference process earlier. Our contributions to fill this gap are twofold: (i) an adaptation of the normalized entropy (NE)-based exit criterion and (ii) a region-based exit criterion that compares segmentation from consecutive early exits. Our analyses reveal that the region-based approach better suits semantic segmentation because it exploits the intermediate early exits more efficiently. Our experiments show that a region-based EE-DNN delivers the same mean intersection over union as an EE-DNN with NE, saving at least 630 million floating-point operations per image on average. |
| Document Type: | Article Conference object |
| File Description: | application/pdf |
| DOI: | 10.1109/icmlcn64995.2025.11140234 |
| Access URL: | https://hal.science/hal-04963398v1 |
| Rights: | STM Policy #29 |
| Accession Number: | edsair.doi.dedup.....fb8b1b78eccf5c20eda5c4f1f6e16f6c |
| Database: | OpenAIRE |
Nájsť tento článok vo Web of Science | Abstract: | Early-exit deep neural networks (EE-DNNs) are multi-output DNNs designed for resource-constrained and latency-sensitive implementations, leveraging auxiliary output layers to partition processing between local, edge, and cloud devices. In the case of semantic segmentation, the literature lacks an efficient exit policy to stop the inference process earlier. Our contributions to fill this gap are twofold: (i) an adaptation of the normalized entropy (NE)-based exit criterion and (ii) a region-based exit criterion that compares segmentation from consecutive early exits. Our analyses reveal that the region-based approach better suits semantic segmentation because it exploits the intermediate early exits more efficiently. Our experiments show that a region-based EE-DNN delivers the same mean intersection over union as an EE-DNN with NE, saving at least 630 million floating-point operations per image on average. |
|---|---|
| DOI: | 10.1109/icmlcn64995.2025.11140234 |