ECCV 2020 Daily - Monday

Why is translation equivariance important? If the model sees chairs, for example, in one corner of the scene during training, it should still be able to reconstruct chairs elsewhere. This is not particularly relevant for single-object scenes, but it is especially important for multi-object scenes. The third technical element is a reconstruction loss tailored to capture overall object geometry . “Most space in the 3D output volume is empty,” Stefan explains. “There is a lot of sparsity . There is a lot of empty space in a scene. Like maybe 90 or 95 per cent of the volume is just empty. We experimented with the focal loss to compensate for this sparsity. It works well, but we also came up with our own loss which is based onthe intersection-over-union metric . We extended it to support continuous values and added support for multiple classes.” 2 Oral Presentation 10 DAILY M o n d a y “Then we built our custom decoder out of this, which starts from the analysis of the image and then expands it with transposed 3D convolutions.” "Building on previous work that reconstructed scenes with a single object, the team proposed three technical extensions."