princeton-vl / DROID-SLAM

BSD 3-Clause "New" or "Revised" License
1.75k stars 295 forks source link

Training for 250000 steps as default on TartanAir, but the result is not good as provided checkpoint #116

Open zhujun3753 opened 10 months ago

zhujun3753 commented 10 months ago

After I trained the model for 250,000 steps, I tested the trained model on the EuRoC dataset MH_01_easy and the results are as follows: max 0.660959 mean 0.014034 median 0.011158 min 0.000964 rmse 0.033552 sse 4.095354 std 0.030476 However, the provided model had following results: max 0.032710 mean 0.010504 median 0.009865 min 0.001174 rmse 0.012018 sse 0.525447 std 0.005840

Why is the gap between RMSE so large?