Unveiling CM-Det: leveraging ConvMixer architecture for advanced object detection

There has recently been a great deal of interest in the application of Vision Transformers (ViT) on complex vision-based tasks, particularly in object detection and tracking. However, the ViT backbone is extremely data hungry and requires large amounts of data and computational power to produce the significant results on the various object detection systems. We present a novel architecture CM-Det, a ConvMixer based Faster RCNN detection system, and compare its performance with the ViT based detection architecture, ViTDet, on the recently published Indian on-road pedestrian detection dataset. We observe that even with the limited computational power, the small size of the training set and no optimization of the training parameters for the ConvMixer base model, CM-Det performs at par with most of the ViTDet variants.