Selection of scene variation possibilities to improve larger-scale data augmentation using Unreal Engine and Synavis. The list is not exhaustive, but should give a good impression that stochastic augmentation should be applied wherever possible.
| Object . | Property . | Description . |
|---|---|---|
| Sun | Position | 2 DoF |
| Sun | Strength | |
| Sun | Ambient light | Intensity |
| Field | Plant position | Varying position 2 DoF, dependent on scene and placing, interior experiment scenes have less potential variability here |
| Field | Plant density | Handled via position, this would be a good way of targeting different DL models that also react sensitively towards changes in complexity |
| Plant | Plant age | determined by parametrization, usually within |
| Plant | Plant leaf bending | Partly visual feature—leaf bending is influenced by environmental conditions. When rendering a wet scene, leaf bending should be increased. For accuracy, it should be determined how large the impact of additional weight is on the leaf structure. If it can be assured that the leaf bending is not taken into account to solve the target problem, this can be estimated. |
| Room | Room lights and diffusion | For digitized experiments, lights should be calibrated to actual brightness values as stated by the bulb manufacturer. |
| Room | Room wall material | RGB texture , Specularity, Roughness |
| Camera | Camera lens properties | A collection of scalars describing the cameras lense and depth-of-field effects in UE |
| Camera | Camera position | Camera paths throughout the scene might further reveal critical insight into algorithm performance in specific and extraordinary circumstances. |
| Post processing | Colour and screen effects | Film grain, movement effects, and more effects can be introduced using Post Process Materials. Not that encoding effects would be captures using the style of encoding and if finetuning towards those effects should be done, the reference should produce effects that might be affected by the encoding the strongest. |
| Object . | Property . | Description . |
|---|---|---|
| Sun | Position | 2 DoF |
| Sun | Strength | |
| Sun | Ambient light | Intensity |
| Field | Plant position | Varying position 2 DoF, dependent on scene and placing, interior experiment scenes have less potential variability here |
| Field | Plant density | Handled via position, this would be a good way of targeting different DL models that also react sensitively towards changes in complexity |
| Plant | Plant age | determined by parametrization, usually within |
| Plant | Plant leaf bending | Partly visual feature—leaf bending is influenced by environmental conditions. When rendering a wet scene, leaf bending should be increased. For accuracy, it should be determined how large the impact of additional weight is on the leaf structure. If it can be assured that the leaf bending is not taken into account to solve the target problem, this can be estimated. |
| Room | Room lights and diffusion | For digitized experiments, lights should be calibrated to actual brightness values as stated by the bulb manufacturer. |
| Room | Room wall material | RGB texture , Specularity, Roughness |
| Camera | Camera lens properties | A collection of scalars describing the cameras lense and depth-of-field effects in UE |
| Camera | Camera position | Camera paths throughout the scene might further reveal critical insight into algorithm performance in specific and extraordinary circumstances. |
| Post processing | Colour and screen effects | Film grain, movement effects, and more effects can be introduced using Post Process Materials. Not that encoding effects would be captures using the style of encoding and if finetuning towards those effects should be done, the reference should produce effects that might be affected by the encoding the strongest. |
Selection of scene variation possibilities to improve larger-scale data augmentation using Unreal Engine and Synavis. The list is not exhaustive, but should give a good impression that stochastic augmentation should be applied wherever possible.
| Object . | Property . | Description . |
|---|---|---|
| Sun | Position | 2 DoF |
| Sun | Strength | |
| Sun | Ambient light | Intensity |
| Field | Plant position | Varying position 2 DoF, dependent on scene and placing, interior experiment scenes have less potential variability here |
| Field | Plant density | Handled via position, this would be a good way of targeting different DL models that also react sensitively towards changes in complexity |
| Plant | Plant age | determined by parametrization, usually within |
| Plant | Plant leaf bending | Partly visual feature—leaf bending is influenced by environmental conditions. When rendering a wet scene, leaf bending should be increased. For accuracy, it should be determined how large the impact of additional weight is on the leaf structure. If it can be assured that the leaf bending is not taken into account to solve the target problem, this can be estimated. |
| Room | Room lights and diffusion | For digitized experiments, lights should be calibrated to actual brightness values as stated by the bulb manufacturer. |
| Room | Room wall material | RGB texture , Specularity, Roughness |
| Camera | Camera lens properties | A collection of scalars describing the cameras lense and depth-of-field effects in UE |
| Camera | Camera position | Camera paths throughout the scene might further reveal critical insight into algorithm performance in specific and extraordinary circumstances. |
| Post processing | Colour and screen effects | Film grain, movement effects, and more effects can be introduced using Post Process Materials. Not that encoding effects would be captures using the style of encoding and if finetuning towards those effects should be done, the reference should produce effects that might be affected by the encoding the strongest. |
| Object . | Property . | Description . |
|---|---|---|
| Sun | Position | 2 DoF |
| Sun | Strength | |
| Sun | Ambient light | Intensity |
| Field | Plant position | Varying position 2 DoF, dependent on scene and placing, interior experiment scenes have less potential variability here |
| Field | Plant density | Handled via position, this would be a good way of targeting different DL models that also react sensitively towards changes in complexity |
| Plant | Plant age | determined by parametrization, usually within |
| Plant | Plant leaf bending | Partly visual feature—leaf bending is influenced by environmental conditions. When rendering a wet scene, leaf bending should be increased. For accuracy, it should be determined how large the impact of additional weight is on the leaf structure. If it can be assured that the leaf bending is not taken into account to solve the target problem, this can be estimated. |
| Room | Room lights and diffusion | For digitized experiments, lights should be calibrated to actual brightness values as stated by the bulb manufacturer. |
| Room | Room wall material | RGB texture , Specularity, Roughness |
| Camera | Camera lens properties | A collection of scalars describing the cameras lense and depth-of-field effects in UE |
| Camera | Camera position | Camera paths throughout the scene might further reveal critical insight into algorithm performance in specific and extraordinary circumstances. |
| Post processing | Colour and screen effects | Film grain, movement effects, and more effects can be introduced using Post Process Materials. Not that encoding effects would be captures using the style of encoding and if finetuning towards those effects should be done, the reference should produce effects that might be affected by the encoding the strongest. |
This PDF is available to Subscribers Only
View Article Abstract & Purchase OptionsFor full access to this pdf, sign in to an existing account, or purchase an annual subscription.
