Depth Estimation
nuscenes

You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

confhead

Per-pixel confidence heads for a frozen Depth Anything V2 on nuScenes, trained by confhead. A head predicts p_error, the probability that the depth of a grid cell is wrong against lidar. Each folder is one training run:

File What it holds
best.pt the head of the epoch with the lowest dev loss
last.pt the head of the last epoch, with the optimizer and the schedule
config.json every setting, the dev scenes, the class weight, the cache fingerprint, the git commit, the run fingerprint
history.json the train and dev loss of each epoch
steps.json the mean train loss five times per epoch, in runs that have it

The heads train on labels from nuScenes, which is under CC BY-NC-SA 4.0 and the nuScenes terms of use. The same terms apply here.

Load a Head

hf download csci566-depthconf/confhead <run>/best.pt --local-dir .
from confhead import load_head
head, inputs, fields = load_head("<run>/best.pt")

Training

confhead/src/confhead/train.py trains every head with AdamW and a cosine learning rate schedule over all steps, against a binary cross-entropy loss on the cells that carry a lidar label, with the class weight below on the positive class. A cell is positive when the absolute relative error of its depth is above tau. The sections below come from the config.json and history.json of each run. Two runs with the same run fingerprint have the same settings, cache build, train images, and dev scenes.

Runs

Run Group Cache Inputs Epochs Best epoch Best dev loss
m1-depth-only inputs nuscenes-trainval depth (1 channels) 8 of 8 8 0.6368
m1-features-only inputs nuscenes-trainval features (64 channels) 8 of 8 8 0.5683
m1-no-depth inputs nuscenes-trainval features, temporal, radar (67 channels) 8 of 8 8 0.5607
m1-no-features inputs nuscenes-trainval depth, temporal, radar (4 channels) 8 of 8 8 0.6129
m1-no-temporal inputs nuscenes-trainval features, depth, radar (67 channels) 8 of 8 8 0.5648
m1-seed1 seed nuscenes-trainval features, depth, temporal, radar (68 channels) 8 of 8 8 0.5622
m1-temporal-only inputs nuscenes-trainval temporal (1 channels) 8 of 8 8 0.6713
m2-radar-lidar-fit radar nuscenes-trainval features, depth, temporal, radar (68 channels) 8 of 8 8 0.5534
m2-radius0 radar nuscenes-trainval features, depth, temporal, radar (68 channels) 8 of 8 8 0.5591
m2-radius2 radar nuscenes-trainval features, depth, temporal, radar (68 channels) 8 of 8 8 0.5581
m2-sweeps1 radar nuscenes-trainval features, depth, temporal, radar (68 channels) 8 of 8 8 0.5582
m3-tau005 labels nuscenes-trainval features, depth, temporal, radar (68 channels) 8 of 8 8 0.4362
m3-tau02 labels nuscenes-trainval features, depth, temporal, radar (68 channels) 8 of 8 8 0.6349
m6-maxdepth50 labels nuscenes-trainval features, depth, temporal, radar (68 channels) 8 of 8 8 0.5633
m6-quarter-data data nuscenes-trainval features, depth, temporal, radar (68 channels) 8 of 8 7 0.5664
run-v0-no-radar inputs nuscenes-trainval features, depth, temporal (66 channels) 8 of 8 8 0.5580
run-v1-radar baseline nuscenes-trainval features, depth, temporal, radar (68 channels) 8 of 8 7 0.5589
run-v1-radar-repeat baseline nuscenes-trainval features, depth, temporal, radar (68 channels) 8 of 8 8 0.5581
trial-run-v1-radar trial nuscenes-trainval features, depth, temporal, radar (68 channels) 1 of 1 1 0.5980

m1-depth-only

Best epoch 8 of 8, dev loss 0.6368. Training took 2.5 hours.

Setting Value
Run fingerprint 1f83d8f4a57dd0de
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs depth (1 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.6635 0.6567 9.62e-04 15.8
2 0.6488 0.6527 8.54e-04 15.5
3 0.6429 0.6493 6.91e-04 16.4
4 0.6383 0.6440 5.00e-04 26.7
5 0.6347 0.6413 3.09e-04 17.1
6 0.6315 0.6394 1.46e-04 14.1
7 0.6289 0.6372 3.81e-05 21.7
8 (best) 0.6276 0.6368 0.00e+00 20.4

m1-features-only

Best epoch 8 of 8, dev loss 0.5683. Training took 2.3 hours.

Setting Value
Run fingerprint 72b701217ec49cda
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features (64 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5852 0.5837 9.62e-04 14.7
2 0.5704 0.5791 8.54e-04 12.8
3 0.5644 0.5735 6.91e-04 15.9
4 0.5601 0.5730 5.00e-04 16.8
5 0.5566 0.5746 3.09e-04 15.5
6 0.5538 0.5706 1.46e-04 16.7
7 0.5517 0.5685 3.81e-05 24.1
8 (best) 0.5506 0.5683 0.00e+00 21.4

m1-no-depth

Best epoch 8 of 8, dev loss 0.5607. Training took 1.9 hours.

Setting Value
Run fingerprint 5cc202ddf3ce2cc4
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, temporal, radar (67 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar radar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5789 0.5776 9.62e-04 19.1
2 0.5629 0.5718 8.54e-04 10.7
3 0.5564 0.5666 6.91e-04 11.7
4 0.5518 0.5657 5.00e-04 13.9
5 0.5481 0.5653 3.09e-04 16.6
6 0.5449 0.5631 1.46e-04 11.9
7 0.5427 0.5609 3.81e-05 12.2
8 (best) 0.5416 0.5607 0.00e+00 16.0

m1-no-features

Best epoch 8 of 8, dev loss 0.6129. Training took 1.7 hours.

Setting Value
Run fingerprint cdccf5cf47e7c6aa
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs depth, temporal, radar (4 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar radar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.6348 0.6304 9.62e-04 11.9
2 0.6195 0.6239 8.54e-04 10.7
3 0.6147 0.6210 6.91e-04 11.8
4 0.6113 0.6185 5.00e-04 13.9
5 0.6083 0.6172 3.09e-04 16.7
6 0.6055 0.6159 1.46e-04 11.8
7 0.6033 0.6133 3.81e-05 12.2
8 (best) 0.6021 0.6129 0.00e+00 16.0

m1-no-temporal

Best epoch 8 of 8, dev loss 0.5648. Training took 1.8 hours.

Setting Value
Run fingerprint beaf0364632ff139
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, radar (67 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar radar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5830 0.5801 9.62e-04 17.1
2 0.5681 0.5775 8.54e-04 10.6
3 0.5622 0.5717 6.91e-04 11.8
4 0.5579 0.5700 5.00e-04 13.9
5 0.5543 0.5700 3.09e-04 16.6
6 0.5513 0.5676 1.46e-04 11.9
7 0.5491 0.5650 3.81e-05 12.2
8 (best) 0.5481 0.5648 0.00e+00 15.9

m1-seed1

Best epoch 8 of 8, dev loss 0.5622. Training took 2.9 hours.

Setting Value
Run fingerprint 9a1d4e3e3756e840
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar radar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,136 of 152,136
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 36.7% of labeled cells
Class weight 1.726
Stats sample 256 images
Seed 1
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5891 0.5815 9.62e-04 19.6
2 0.5732 0.5807 8.54e-04 18.6
3 0.5668 0.5722 6.91e-04 24.2
4 0.5621 0.5687 5.00e-04 25.9
5 0.5582 0.5655 3.09e-04 17.8
6 0.5550 0.5633 1.46e-04 25.3
7 0.5526 0.5623 3.81e-05 24.0
8 (best) 0.5515 0.5622 0.00e+00 21.1

m1-temporal-only

Best epoch 8 of 8, dev loss 0.6713. Training took 2.2 hours.

Setting Value
Run fingerprint 61bb97acc13c1ff5
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs temporal (1 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.7036 0.7078 9.62e-04 11.1
2 0.6863 0.6868 8.54e-04 10.8
3 0.6750 0.6836 6.91e-04 15.9
4 0.6704 0.6779 5.00e-04 16.8
5 0.6669 0.6780 3.09e-04 15.4
6 0.6636 0.6754 1.46e-04 16.7
7 0.6615 0.6719 3.81e-05 24.1
8 (best) 0.6604 0.6713 0.00e+00 21.3

m2-radar-lidar-fit

Best epoch 8 of 8, dev loss 0.5534. Training took 2.8 hours.

Setting Value
Run fingerprint d4499db53fc710cc
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar lidar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5731 0.5722 9.62e-04 18.5
2 0.5564 0.5659 8.54e-04 26.5
3 0.5499 0.5610 6.91e-04 20.8
4 0.5452 0.5591 5.00e-04 15.7
5 0.5414 0.5586 3.09e-04 23.6
6 0.5380 0.5568 1.46e-04 22.2
7 0.5357 0.5538 3.81e-05 19.3
8 (best) 0.5345 0.5534 0.00e+00 21.1

m2-radius0

Best epoch 8 of 8, dev loss 0.5591. Training took 2.7 hours.

Setting Value
Run fingerprint df988ef5e3f79705
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar radar-aligned, radius 0 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5778 0.5778 9.62e-04 16.1
2 0.5619 0.5716 8.54e-04 25.8
3 0.5553 0.5662 6.91e-04 20.9
4 0.5505 0.5645 5.00e-04 15.8
5 0.5467 0.5632 3.09e-04 23.6
6 0.5435 0.5623 1.46e-04 22.2
7 0.5413 0.5593 3.81e-05 19.3
8 (best) 0.5401 0.5591 0.00e+00 21.1

m2-radius2

Best epoch 8 of 8, dev loss 0.5581. Training took 2.7 hours.

Setting Value
Run fingerprint 70165d47780ccfc6
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar radar-aligned, radius 2 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5774 0.5775 9.62e-04 20.2
2 0.5615 0.5712 8.54e-04 18.5
3 0.5549 0.5653 6.91e-04 24.3
4 0.5502 0.5630 5.00e-04 21.3
5 0.5463 0.5635 3.09e-04 20.4
6 0.5430 0.5611 1.46e-04 23.6
7 0.5407 0.5582 3.81e-05 15.8
8 (best) 0.5396 0.5581 0.00e+00 18.9

m2-sweeps1

Best epoch 8 of 8, dev loss 0.5582. Training took 2.7 hours.

Setting Value
Run fingerprint a1fb60ce9682d89c
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar radar-aligned, radius 1 grid cells, sweeps: 1, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5777 0.5773 9.62e-04 17.6
2 0.5620 0.5715 8.54e-04 18.1
3 0.5556 0.5659 6.91e-04 24.3
4 0.5509 0.5638 5.00e-04 21.3
5 0.5471 0.5624 3.09e-04 20.4
6 0.5438 0.5608 1.46e-04 23.6
7 0.5416 0.5585 3.81e-05 15.8
8 (best) 0.5404 0.5582 0.00e+00 18.9

m3-tau005

Best epoch 8 of 8, dev loss 0.4362. Training took 2.7 hours.

Setting Value
Run fingerprint 7edeeffd26dd0a01
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.05, depth 1 to 80 m
Radar radar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 55.6% of labeled cells
Class weight 0.798
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.4536 0.4464 9.62e-04 16.4
2 0.4441 0.4430 8.54e-04 24.8
3 0.4401 0.4404 6.91e-04 23.4
4 0.4372 0.4395 5.00e-04 20.8
5 0.4349 0.4388 3.09e-04 24.7
6 0.4328 0.4377 1.46e-04 17.7
7 0.4315 0.4364 3.81e-05 18.7
8 (best) 0.4307 0.4362 0.00e+00 14.2

m3-tau02

Best epoch 8 of 8, dev loss 0.6349. Training took 2.8 hours.

Setting Value
Run fingerprint 26268abba55c657f
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.2, depth 1 to 80 m
Radar radar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 21.4% of labeled cells
Class weight 3.684
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.6615 0.6728 9.62e-04 18.5
2 0.6382 0.6531 8.54e-04 22.4
3 0.6293 0.6471 6.91e-04 23.1
4 0.6224 0.6441 5.00e-04 22.0
5 0.6170 0.6386 3.09e-04 21.9
6 0.6126 0.6372 1.46e-04 21.7
7 0.6093 0.6350 3.81e-05 20.8
8 (best) 0.6077 0.6349 0.00e+00 19.5

m6-maxdepth50

Best epoch 8 of 8, dev loss 0.5633. Training took 2.8 hours.

Setting Value
Run fingerprint 7a7bdf1005dda54c
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 50 m
Radar radar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.4% of labeled cells
Class weight 1.675
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5822 0.5809 9.62e-04 16.4
2 0.5661 0.5757 8.54e-04 22.4
3 0.5595 0.5708 6.91e-04 23.1
4 0.5547 0.5684 5.00e-04 22.0
5 0.5509 0.5682 3.09e-04 21.9
6 0.5476 0.5662 1.46e-04 21.7
7 0.5454 0.5633 3.81e-05 20.8
8 (best) 0.5442 0.5633 0.00e+00 19.5

m6-quarter-data

Best epoch 7 of 8, dev loss 0.5664. Training took 0.8 hours.

Setting Value
Run fingerprint e1b0512a276adc6f
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar radar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 38,000 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5932 0.5966 9.62e-04 4.1
2 0.5778 0.5816 8.54e-04 4.9
3 0.5698 0.5845 6.91e-04 5.3
4 0.5649 0.5750 5.00e-04 6.6
5 0.5609 0.5724 3.09e-04 6.9
6 0.5567 0.5689 1.46e-04 6.1
7 (best) 0.5535 0.5664 3.81e-05 6.5
8 0.5531 0.5667 0.00e+00 6.2

run-v0-no-radar

Best epoch 8 of 8, dev loss 0.5580. Training took 2.0 hours.

Setting Value
Run fingerprint 6d333621de84321f
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal (66 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5777 0.5761 9.62e-04 21.7
2 0.5614 0.5710 8.54e-04 14.9
3 0.5550 0.5652 6.91e-04 14.6
4 0.5502 0.5631 5.00e-04 13.9
5 0.5464 0.5625 3.09e-04 16.6
6 0.5431 0.5608 1.46e-04 11.9
7 0.5409 0.5582 3.81e-05 12.2
8 (best) 0.5397 0.5580 0.00e+00 16.0

run-v1-radar

Best epoch 7 of 8, dev loss 0.5589. Training took 3.1 hours.

Setting Value
Run fingerprint eb6263479d436417
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar radar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5776 0.5760 9.62e-04 41.1
2 0.5617 0.5704 8.54e-04 26.5
3 0.5552 0.5664 6.91e-04 29.8
4 0.5504 0.5647 5.00e-04 28.3
5 0.5466 0.5631 3.09e-04 16.8
6 0.5434 0.5618 1.46e-04 13.0
7 (best) 0.5411 0.5589 3.81e-05 14.5
8 0.5399 0.5589 0.00e+00 14.8

run-v1-radar-repeat

Best epoch 8 of 8, dev loss 0.5581. Training took 2.8 hours.

Setting Value
Run fingerprint eb6263479d436417
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar radar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 8
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 152,124 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 0.5777 0.5764 9.62e-04 26.6
2 0.5617 0.5696 8.54e-04 25.8
3 0.5553 0.5656 6.91e-04 29.8
4 0.5506 0.5644 5.00e-04 28.4
5 0.5468 0.5635 3.09e-04 16.8
6 0.5436 0.5612 1.46e-04 13.0
7 0.5414 0.5583 3.81e-05 14.6
8 (best) 0.5402 0.5581 0.00e+00 14.8

trial-run-v1-radar

Best epoch 1 of 1, dev loss 0.5980. Training took 0.1 hours.

Setting Value
Run fingerprint 910b0253ded4d0d3
Cache nuscenes-trainval, fingerprint 99d44ff8c71c5355, grid 132 ร— 74
Depth model Depth Anything V2 vitb, short side 518, tap output_conv1
Inputs features, depth, temporal, radar (68 channels)
Labels lidar-aligned, tau 0.1, depth 1 to 80 m
Radar radar-aligned, radius 1 grid cells, sweeps: every cached sweep, depth 1 to 80 m
Head widths [64, 64, 32]
Epochs 1
Batch size 8
Learning rate 0.001
Weight decay 0.0001
Train images per epoch 10,000 of 152,124
Dev set 69 scenes, 3,000 images (dev_fraction 0.1, dev_limit 3,000)
Positive rate 37.6% of labeled cells
Class weight 1.660
Stats sample 256 images
Seed 0
Software git 022e800, torch 2.14.1+cu126, device cuda
Epoch Train loss Dev loss Learning rate at end Minutes
1 (best) 0.6073 0.5980 0.00e+00 3.3
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support