Really great work! A little question, I saw the teaser.png (https://github.com/robot-learning-freiburg/PanopticBEV/blob/master/images/teaser.png) was generated from 6 cameras (should be nuscenes dataset?), but in nuscenes.ini, only listed cam_front. [cameras] intrinsics = {"fx": 1266.4172, "fy": 51276.4172, "px": 816.267, "py": 491.507} extrinsics = {"translation": (0.0, 0.6, 1.85), "rotation": (-90, 0, 180)} bev_params = {"f": 336, "cam_z": 26}

So how can I enable multi-camera input for PanopticBEV? Or can you post a multi-camera version?

Dear authors,

thanks for your great work. I have trained one model myself and want to see the visualized output(like Fig.1 in your paper). I found one script under the path panoptic_bev/utils/visualization.py. But don't know the way to use it.

Could please tell me the way to visualize the result?

Bests,

Hi, Have you seen this error before? I tried training from scratch and get this for:

for it, sample in enumerate(dataloader): sample = {k: sample[k].cuda(device=varargs['device'], non_blocking=True) for k in NETWORK_INPUTS}

Thanks, Shubhankar

I downloaded nuscenes_panopticbev.zip, and found only front camera ground truth, but I need labels for all 6 cameras, so can you release the dataset annotation tool? Or the whole annotated nuscenes for 6 cameras, that will be helpful.

For KITTI-360 data, What is the starting coordinate of the bev map? I mean the bev map's left border coordinate, The bev maps contain real-world distance relationships, so the start coordinate is very import, but it was not given in your paper. There is also another question, Are the bev maps in the LiDAR coordinate system or the camera coordinate system?

Which data set of Nuscenes are you using? How do you set these Settings? I would appreciate it if you could give me a reply!

Originally posted by @laozheng1 in https://github.com/robot-learning-freiburg/PanopticBEV/issues/10#issuecomment-1072342490

I change the "scale=1.0" to "scale=0.5" in nuscenes.ini, and re-trained the model, expecting similar results, but the object-detection perform very bad in the smaller scale.

result from "scale=1.0", fairly good.

but result from "scale=0.5", nothing detected !

how should I tune the model after changing the scale, any suggestion?

Dear all, we trained the model on Nuscenes and Kitti360 respectively. The Nuscenes work fine. But for Kitti360, the results don't seem very well. We attached the result at the end. I didn't change any code or configuration. I found that the size of the FV image in Kitti360 is 768 * 1400 on paper but 384 * 1400 in the configuration file in code. Which one is correct?

01:16:57 ---------------- Semantic mIoU ---------------- 01:16:57 road : 0.66676 01:16:57 sidewalk : 0.29211 01:16:57 building : 0.26043 01:16:57 wall : 0.03577 01:16:57 vegetation : 0.34960 01:16:57 terrain : 0.14224 01:16:57 occlusion : 0.40774 01:16:57 person : 0.00000 01:16:57 rider : 0.00789 01:16:57 car : 0.35977 01:16:57 truck : 0.07531 01:16:57 ---------------- Panoptic Scores ---------------- 01:16:57 po_miou : 0.23538 01:16:57 sem_miou : 0.23615 01:16:57 pq : 0.14057 01:16:57 pq_stuff : 0.17719 01:16:57 pq_thing : 0.07647 01:16:57 sq : 0.55955 01:16:57 sq_stuff : 0.60958 01:16:57 sq_thing : 0.47198 01:16:57 rq : 0.21159 01:16:57 rq_stuff : 0.26759