benallaldev/carddd

Car Damage Detection & Segmentation

Model: Mask2Former + Swin-L backbone Task: Instance segmentation of car damage Framework: Detectron2

Classes

ID	Label	Description
0	rach	Scratch
1	mat_bo_phan	Missing part
2	be_den	Broken headlight
3	thung	Hole / puncture
4	tray_son	Paint damage
5	mop_lom	Dent
6	vo_kinh	Broken glass

Training Dataset

• Train images: 11,621
• Val images: 2,324
• Annotation format: Polygon instance masks

Latest Checkpoint

• File: checkpoint_epoch50_iter145300.pth (also available as model_latest.pth)
• Iteration: 145300
• Epoch: 50.0
• Mask AP: 30.8

Evaluation Results (Validation Set)

Metric	Value
Mask AP (all IoU)	30.8
Mask AP50	52.4
Mask AP75	28.3

Per-class Mask AP:

Class	AP
rach	16.0
mat_bo_phan	51.5
be_den	31.0
thung	30.3
tray_son	12.2
mop_lom	15.1
vo_kinh	59.3

Training Progress

Loss (sampled):

Iteration	Total Loss
52319	13.9277
61579	14.3633
70919	13.3874
80259	13.6256
89599	13.0817
98939	11.2451
108279	11.0778
117439	8.3518
126779	9.4671
135879	9.0137
145219	8.6403
145299	9.7138

Mask AP over time:

Iteration	AP	AP50
58119	31.0	53.1
116239	30.8	52.4

---

Last updated: 2026-04-10 02:39 UTC Pushed automatically during training.