MAMBA: Multi-level Aggregation via Memory Bank for Video Object Detection

State-of-the-art video object detection methods maintain a memory structure, either sliding window or queue, to enhance the current frame using attention mechanisms. However, we argue that these structures are not efficient sufficient because of two implied operations: (1) concatenating all features in for enhancement, leading heavy computational cost; (2) frame-wise updating, preventing from capturing more temporal information. In this paper, propose multi-level aggregation architecture via bank called MAMBA. Specifically, our employs novel operations eliminate disadvantages existing methods: light-weight key-set construction which can significantly reduce fine-grained feature-wise updating strategy enables method utilize knowledge whole video. To better complementary levels, i.e., feature maps and proposals, further generalized enhancement operation (GEO) aggregate unified manner. We conduct extensive evaluations on challenging ImageNetVID dataset. Compared with state-of-the-art methods, achieves superior performance terms both speed accuracy. More remarkably, MAMBA mAP 83.7%/84.6% at 12.6/9.1 FPS ResNet-101.