We have used Single Shot Detector SSD using MobileNet as the feature extractor (which was trained using ImageNet) and pre-trained on COCO dataset as our baseline.
CPU: Intel(R) Core(TM) i7-7820X CPU @ 3.60GHz
GPU0: Nvidia TITAN Xp
GPU1: Nvidia TITAN Xp
RAM: 64GB
OS: Ubuntu 18.04 Bionic
The network is automatically split between the 2 GPUs. Our setup allowed us to use a batch size of 8 in all training scenarios.
The following table shows which dataset is used in each run, and its composition.
| Run | Training Set | # sim images | # real images | SSD Checkpoint | Epochs |
|---|---|---|---|---|---|
| fine_tune/real | real.record | 0 | 175 | COCO | 16k |
| sim_no_camera | sim_no_camera.record | 30k | 0 | - | ~90 |
| sim_big_camera | sim_big_camera.record | 30k | 0 | - | ~90 |
| fine_tune/sim_no_camera | real.record | 0 | 175 | sim_no_camera | ~2.2k |
| fine_tune/sim_big_camera | real.record | 0 | 175 | sim_big_camera | ~2.2k |
| sim_no_camera_* | sim_no_camera_*.record | 6k | 0 | - | ~130 |
| fine_tune/sim_no_camera_* | real.record | 0 | 175 | sim_no_camera_* | ~1.1k |
The baseline is fine_tune/real which corresponds to using SSD pre-trained on COCO and fine-tuning on real image dataset.
The asterisk (*) in sim_no_camera_* represents each of the 5 subdatasets, namely base,flat,chess,gradient,perlin.
We used exponentially decaying learning rate, α = α_0 * k ^ floor(- iteration / t), with k = 0.95.
| Run | Learning rate α_0 | Decay steps t | Iterations |
|---|---|---|---|
| fine_tune/real | 0.004 | 100k | 350k |
| sim_no_camera | 0.008 | 50k | 350k |
| sim_big_camera | 0.004 | 50k | 350k |
| fine_tune/sim_no_camera | 0.001 | - | 50k |
| fine_tune/sim_big_camera | 0.001 | - | 50k |
| sim_no_camera_* | 0.004 | 50k | 100k |
| fine_tune/sim_no_camera_* | 0.001 | - | 25k |