Fault resilience in computer vision algorithms is paramount in critical applications such as autonomous driving or surveillance. Convolutional neural networks (CNNs) are usually used in these tasks to identify objects of interest, which are passed to other decisional algorithms and used to take specific actions. However, incorrect detections due to computation errors could pose a safety risk. In this work, we present ERODE (Error Resilient Object DetEction), a framework that can be paired with a CNN to filter the detections by identifying possible errors and restoring the correct predictions, improving the faultresilience of the system. The proposed framework leverages the temporal correlation among consecutive images and CNN outputs, using motion estimation and tracking techniques to infer whether computation errors have occurred and, in that case, produce a new set of outputs. In order to evaluate the performance, precision and recall of the CNN with and without ERODE support have been computed and compared using the MOT17DET dataset, and EfficientDet D0 quantized to 16-bit, with errors injected in the activations computed during the inference. The experimental results show significantly reduced task accuracy degradation induced by bit-flips, proving that ERODE can increase the system’s fault resilience.
ERODE: Error Resilient Object DetEction by Recovering Bounding Box and Class Information / Valpreda, Emanuele; Palumbo, Giuseppe; Caon, Michele; Masera, Guido; Martina, Maurizio. - ELETTRONICO. - (2023), pp. 277-280. (Intervento presentato al convegno 18th International Conference on PhD Research in Microelectronics and Electronics tenutosi a Valencia (Spain) nel 18-21 June 2023) [10.1109/PRIME58259.2023.10161894].
ERODE: Error Resilient Object DetEction by Recovering Bounding Box and Class Information
Valpreda, Emanuele;Caon, Michele;Masera, Guido;Martina, Maurizio
2023
Abstract
Fault resilience in computer vision algorithms is paramount in critical applications such as autonomous driving or surveillance. Convolutional neural networks (CNNs) are usually used in these tasks to identify objects of interest, which are passed to other decisional algorithms and used to take specific actions. However, incorrect detections due to computation errors could pose a safety risk. In this work, we present ERODE (Error Resilient Object DetEction), a framework that can be paired with a CNN to filter the detections by identifying possible errors and restoring the correct predictions, improving the faultresilience of the system. The proposed framework leverages the temporal correlation among consecutive images and CNN outputs, using motion estimation and tracking techniques to infer whether computation errors have occurred and, in that case, produce a new set of outputs. In order to evaluate the performance, precision and recall of the CNN with and without ERODE support have been computed and compared using the MOT17DET dataset, and EfficientDet D0 quantized to 16-bit, with errors injected in the activations computed during the inference. The experimental results show significantly reduced task accuracy degradation induced by bit-flips, proving that ERODE can increase the system’s fault resilience.File | Dimensione | Formato | |
---|---|---|---|
erode_prime2023_final_private.pdf
accesso aperto
Descrizione: Versione post print autore
Tipologia:
2. Post-print / Author's Accepted Manuscript
Licenza:
PUBBLICO - Tutti i diritti riservati
Dimensione
1.04 MB
Formato
Adobe PDF
|
1.04 MB | Adobe PDF | Visualizza/Apri |
Valpreda-Erode.pdf
non disponibili
Tipologia:
2a Post-print versione editoriale / Version of Record
Licenza:
Non Pubblico - Accesso privato/ristretto
Dimensione
1.4 MB
Formato
Adobe PDF
|
1.4 MB | Adobe PDF | Visualizza/Apri Richiedi una copia |
Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/11583/2979532