Question about Background Object Stationarity & Poor Natural Image Inference Quality

[sustecher]

I hope to confirm whether background objects are designed to be recognized as stationary by default. For example, in a road scene featuring a person running, the inference result appears to indicate that the runner is stationary, which seems inconsistent with the actual motion state.

Additionally, I would like to ask if the application scenario of this project is restricted to the Minecraft game environment? I have attempted to conduct inference on natural images, but the quality of the resulting images is unfortunately quite poor.

To help with troubleshooting, I have attached three test examples of my own for your reference.

case1.mp4

test_058_20260330_230555_264326_6879_r1.mp4

test_059_20260330_230831_271277_6402_r2.mp4

[sustecher]

In this test example, the river is identified as stationary, which seems to indicate that background elements are treated as static by default.

However, in the official demos, background elements are capable of being dynamic – for instance, people running, horses moving, and cars driving forward (all appear as dynamic background elements).

Could you kindly clarify the reason for this discrepancy? Is it possible that there is an issue with our inference process, or has the official team not released the weight parameters that enable greater dynamism?

debug_single_current_iteration_0.mp4

test_case_1_actions_grid_ijkl_2x2.mp4

test_case_1_actions_grid_wasd_2x2.mp4