We consider the problem of localizing a spatio-temporal tube in a videocorresponding to a given text query. This is a challenging task that requiresthe joint and efficient modeling of temporal, spatial and multi-modalinteractions. To address this task, we propose TubeDETR, a transformer-basedarchitecture inspired by the recent success of such models for text-conditionedobject detection. Our model notably includes: (i) an efficient video and textencoder that models spatial multi-modal interactions over sparsely sampledframes and (ii) a space-time decoder that jointly performs spatio-temporallocalization. We demonstrate the advantage of our proposed components throughan extensive ablation study. We also evaluate our full approach on thespatio-temporal video grounding task and demonstrate improvements over thestate of the art on the challenging VidSTG and HC-STVG benchmarks. Code andtrained models are publicly available athttps://antoyang.github.io/tubedetr.html.