Learning in an Uncertain World: Representing Ambiguity Through Multiple Hypotheses
Citations
470 citations
Cites methods or result from "Learning in an Uncertain World: Rep..."
...Mirroring an existing setup from [31] and [37], we filter out low-probability trajectories (we set the threshold to 0....
[...]
...To address this issue, we propose to use a novel Multiple-Trajectory Prediction (MTP) loss, motivated by [37], that explicitly models the multimodality of the trajectory space....
[...]
...[37] reported this issue for MDNs and found multi-hypothesis models to be less affected, as confirmed by our results....
[...]
...To overcome this problem researchers proposed training an ensemble of networks [36], or a single network to produce M different outputs for M different hypotheses [37] using a loss that only accounts for the closest prediction to ground truth labels....
[...]
295 citations
Cites background or methods from "Learning in an Uncertain World: Rep..."
...Aiming for diverse semantic segmentation outputs, the works of [12] and [13] propose to branch off M heads after the last layer of a deep net each of which contributes one output variant....
[...]
...This has been explored in [11] and [1] using an ensemble of deep networks, and in [12] and [13] using one common deep network with M heads....
[...]
193 citations
159 citations
134 citations
Cites methods from "Learning in an Uncertain World: Rep..."
...To encourage all members of the ensemble to be trained equally, as mentioned in [43], we made this association soft by routing the gradient to the closest network with a 0....
[...]
References
123,388 citations
73,978 citations
49,914 citations
42,067 citations
33,597 citations
"Learning in an Uncertain World: Rep..." refers methods in this paper
...Additionally, we adapt the concept from [33] to drop out full predictions with some low probability (1% in our experiments)....
[...]