YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information

The work targets a fact ignored by many deep learning methods: as input data undergoes layer-by-layer feature extraction and spatial transformation, large amounts of information are lost, an issue tied to the information bottleneck and reversible functions. To address it, the authors propose Programmable Gradient Information (PGI), which supplies complete input information for the target task so reliable gradient information can be obtained to update network weights, and they design the Generalized Efficient Layer Aggregation Network (GELAN) based on gradient path planning.

Verified on the MS COCO object detection dataset, GELAN uses only conventional convolution operators yet achieves better parameter utilization than state-of-the-art methods built on depth-wise convolution. PGI applies across models from lightweight to large and, by preserving complete information, enables train-from-scratch models to surpass state-of-the-art models pre-trained on large datasets, with source code released publicly.